Open Source as it gets in this space, top notch developer documentation, and prices insanely low, while delivering frontier model capabilities. So basically, this is from hackers to hackers. Loving it!
Also, note that there's zero CUDA dependency. It runs entirely on Huawei chips. In other words, Chinese ecosystem has delivered a complete AI stack. Like it or not, that's a big news. But what's there not to like when monopolies break down?
Just ask it for a summary of the USA’s role in Iran, Gaza, Lebanon and its recent threats against Panama, Cuba and Greenland! It might be able to keep track.
Yeah, it’s an interesting one. I think inertia and expectations at this point? I don’t think the big labs anticipated how low the model switching costs would be and how quickly their leads would be eroded (by each other and the upstarts)
They are developing their moats with the platform tooling around it right now though. Look at Anthropic with Routines and OpenAI with Agents. Drop that capability in to a business with loose controls and suddenly you have a very sticky product with high switching costs. Meanwhile if you stick with purely the ‘chat’ use cases, even Cowork and scheduled tasks, you maintain portability.
If you want other people to know whether you're being genuine or sarcastic, you'll have to put a bit more effort into your comments. Your comment just adds noise.
There are quite a few comments here about benchmark and coding performance. I would like to offer some opinions regarding its capacity for mathematics problems in an active research setting.
I have a collection of novel probability and statistics problems at the masters and PhD level with varying degrees of feasibility. My test suite involves running these problems through first (often with about 2-6 papers for context) and then requesting a rigorous proof as followup. Since the problems are pretty tough, there is no quantitative measure of performance here, I'm just judging based on how useful the output is toward outlining a solution that would hopefully become publishable.
Just prior to this model, Gemini led the pack, with GPT-5 as a close second. No other model came anywhere near these two (no, not even Claude). Gemini would sometimes have incredible insight for some of the harder problems (insightful guesses on relevant procedures are often most useful in research), but both of them tend to struggle with outlining a concrete proof in a single followup prompt. This DeepSeek V4 Pro with max thinking does remarkably well here. I'm not seeing the same level of insights in the first response as Gemini (closer to GPT-5), but it often gets much better in the followup, and the proofs can be _very_ impressive; nearly complete in several cases.
Given that both Gemini and DeepSeek also seem to lead on token performance, I'm guessing that might play a role in their capacity for these types of problems. It's probably more a matter of just how far they can get in a sensible computational budget.
Despite what the benchmarks seem to show, this feels like a huge step up for open-weight models. Bravo to the DeepSeek team!
>we implement end-to-end, bitwise batch-invariant, and deterministic kernels with minimal performance overhead
Pretty cool, I think they're the first to guarantee determinism with the fixed seed or at the temperature 0. Google came close but never guaranteed it AFAIK. DeepSeek show their roots - it may not strictly be a SotA model, but there's a ton of low-level optimizations nobody else pays attention to.
I’d like somebody to explain to me how the endless comments of "bleeding edge labs are subsidizing the inference at an insane rate" make sense in light of a humongous model like v4 pro being $4 per 1M. I’d bet even the subscriptions are profitable, much less the API prices.
API prices may be profitable. Subscriptions may still be subsidized for power users. Free tiers almost certainly are. And frontier labs may be subsidizing overall business growth, training, product features, and peak capacity, even if a normal metered API call is profitable on marginal inference.
This price is high even because of the current shortage of inference cards available to DeepSeek; they claimed in their press release that once the Ascend 950 computing cards are launched in the second half of the year, the price of the Pro version will drop significantly
I was thinking the same. How can it be than other providers can offer third-party open source models with roughly the similar quality like this, Kimi K2.6 or GLM 5.1 for 10 times less the price? How can it be that GPT 5.5 is suddenly twice the price as GPT 5.4 while being faster? I don't believe that it's a bigger, more expensive model to run, it's just they're starting to raise up the prices because they can and their product is good (which is honest as long as they're transparent with it). Honestly the movement about subscription costing the company 20 times more than we're paying is just a PR movement to justify the price hike.
But seriously, it just stems from the fact some people want AI to go away. If you set your conclusion first, you can very easily derive any premise. AI must go away -> AI must be a bad business -> AI must be losing money.
Before the AI bubble that will burst any time now, there was the AI winter that would magically arrive before the models got good enough to rival humans.
My thoughts exactly. I also believe that subscription services are profitable, and the talk about subsidies is just a way to extract higher profit margins from the API prices businesses pay.
I mean, not one "bleeding edge" lab has stated they are profitable. They don't publish financials aside from revenue. And in Anthropic's case, they fuck with pricing every week. Clearly something is wrong here.
As this is a new arch with tons of optimisations, it'll take some time for inference engines to support it properly, and we'll see more 3rd party providers offer it. Once that settles we'll have a median price for an optimised 1.6T model, and can "guesstimate" from there what the big labs can reasonably serve for the same price. But yeah, it's been said for a while that big labs are ok on API costs. The only unknown is if subscriptions were profitable or not. They've all been reducing the limits lately it seems.
> I’d like somebody to explain to me how the endless comments of "bleeding edge labs are subsidizing the inference at an insane rate" make sense in light of a humongous model like v4 pro being $4 per 1M. I’d bet even the subscriptions are profitable, much less the API prices.
One answer - Chinese Communist Party. They are being subsidized by the state.
While SWE-bench Verified is not a perfect benchmark for coding, AFAIK, this is the first open-weights model that has crossed the threshold of 80% score on this by scoring 80.6%.
Back in Nov 2025, Opus 4.5 (80.9%) was the first proprietary model to do so.
Their audience is people who build stuff, techs audience is enterprise CEOs and politicians, and anyone else happy to hype up all the questionably timed releases and warnings of danger, white collar irrelevence, or promises of utopian paradise right before a funding round.
doesn't it get tiring after a while? using the same (perceived) gotcha, over and over again, for three years now?
no one is ever going to release their training data because it contains every copyrighted work in existence. everyone, even the hecking-wholesome safety-first Anthropic, is using copyrighted data without permission to train their models. there you go.
There is an easy fix already in widespread use: "open weights".
It is very much a valuable thing already, no need to taint it with wrong promise.
Though I disagree about being used if it was indeed open source: I might not do it inside my home lab today, but at least Qwen and DeepSeek would use and build on what eg. Facebook was doing with Llama, and they might be pushing the open weights model frontier forward faster.
American companies want a scan of your asshole for the privilege of paying to access their models, and unapologetically admit to storing, analyzing, training on, and freely giving your data to any authorities if requested. Chinese ulteriority is hypothetical, American is blatant.
It’s not remotely hypothetical you’d have to be living under a rock to believe that. And the fusion with a one-party state government that doesn’t tolerate huge swathes of thoughtspace being freely discussed is completely streamlined, not mediated by any guardrails or accountability.
This “no harm to me” meme about a foreign totalitarian government (with plenty of incentive to run influence ops on foreigners) hoovering your data is just so mind-bogglingly naive.
As a non-American, everything you wrote other than "one party" applies to the current US regime.
Relatively speaking, DeepSeek is less untrustworthy than Grok.
When I try ChatGPT on current events from the White House it interprets them as strange hypotheticals rather than news, which is probably more a problem with DC than with GPT, but whatever.
> And the fusion with a one-party state government that doesn’t tolerate huge swathes of thoughtspace being freely discussed
That would be a great argument if the American models weren’t so heavily censored.
The Chinese model might dodge a question if I ask it about 1-2 specific Chinese cultural issues but then it also doesn’t moralize me at every turn because I asked it to use a piece of security software.
>This “no harm to me” meme about a foreign totalitarian government (with plenty of incentive to run influence ops on foreigners) hoovering your data is just so mind-bogglingly naive.
> This “no harm to me” meme about a foreign totalitarian government (with plenty of incentive to run influence ops on foreigners) hoovering your data is just so mind-bogglingly naive.
This is why I’ve been urging everyone I know to move away from American based services and providers. It’s slow but honest work.
Pretty sure you guys have a strong laws about free-speech, and criticizing elites is part of that. Though there are some groups that do not really want the 1st amendment to be a thing.
Foreigners are literally being denied entry into the country due to opposing viewpoints expressed on social media. People have to disable FaceID on their phones prior to going through customs in case an agent decides to investigate whether their political views are in opposition to the current administration.
It's a little sad that tech now comes down to geopolitics, but if you're not in the USA then what is the difference? I'm Danish, would I rather give my data to China or to a country which recently threatened the kingdom I live in with military invasion? Ideally I'd give them to Mistral, but in reality we're probably going to continue building multi-model tools to make sure we share our data with everyone equally.
> Internet comments say that open sourcing is a national strategy, a loss maker subsidized by the government. On the contrary, it is a commercial strategy and the best strategy available in this industry.
This sounds whole lot like potatoh potahto. I think the former argument is very much the correct one: China can undercut everyone and win, even at a loss. Happened with solar panels, steel, evs, sea food - it's a well tested strategy and it works really well despite the many flavors it comes in.
That being said a job well done for the wrong reasons is still a job well done so we should very much welcome these contributions, and maybe it's good to upset western big tech a bit so it's remains competitive.
It is not only that Chinese labs can undercut on price. It is that they must. They must give away their models for free by open sourcing them, and they must even give away free inference services for people to try them. That is the point of the post.
Please don't slander the most open AI company in the world. Even more open than some non-profit labs from universities. DeepSeek is famous for publishing everything. They might take a bit to publish source code but it's almost always there. And their papers are extremely pro-social to help the broader open AI community. This is why they struggle getting funded because investors hate openness. And in China they struggle against the political and hiring power of the big tech companies.
And DeepSeek often has very cool new approaches to AI copied by the rest. Many others copied their tech. And some of those have 10x or 100x the GPU training budget and that's their moat to stay competitive.
It’s not slander to say something true. These are open weights, not open source. They don’t provide the training data or the methodology requires to reproduce these weights.
So you can’t see what facts are pruned out, what biases were applied, etc. Even more importantly, you can’t make a slightly improved version.
This model is as open source as a windows XP installation ISO.
Its on OR - but currently not available on their anthropic endpoint. OR if you read this, pls enable it there! I am using kimi-2.6 with Claude Code, works well, but Deepseek V4 gives an error:
`https://openrouter.ai/api/messages with model=deepseek/deepseek-v4-pro, OR returns
an error because their Anthropic-compat translator doesn't cover V4 yet. The Claude CLI dutifully surfaces that error as "model...does not exist"
I’m deeply interested and invested in the field but I could really use a support group for people burnt out from trying to keep up with everything. I feel like we’ve already long since passed the point where we need AI to help us keep up with advancements in AI.
The players barely ever change. People don't have problems following sports, you shouldn't struggle so much with this once you accept top spot changes.
I didn't express this well but my interest isn't "who is in the top spot", and is more _why and _how various labs get the results they do. This is also magnified by the fact that I'm not only interested in hosted providers of inference but local models as well. What's your take on the best model to run for coding on 24GB of VRAM locally after the last few weeks of releases? Which harness do you prefer? What quants do you think are best? To use your sports metaphor it's more than following the national leagues but also following college and even high school leagues as well. And the real interest isn't even who's doing well but WHY, at each level.
It is funny seeing people ping pong between Anthropic and ChatGPT, with similar rhetoric in both directions.
At this point I would just pick the one who's "ethics" and user experience you prefer. The difference in performance between these releases has had no impact on the meaningful work one can do with them, unless perhaps they are on the fringes in some domain.
Personally I am trying out the open models cloud hosted, since I am not interested in being rug pulled by the big two providers. They have come a long way, and for all the work I actually trust to an LLM they seem to be sufficient.
It honestly has all kinda felt like more of the same ever since maybe GPT4?
New model comes out, has some nice benchmarks, but the subjective experience of actually using it stays the same. Nothing's really blown my mind since.
Feels like the field has stagnated to a point where only the enthusiasts care.
For coding Opus 4.5 in q3 2025 was still the best model I've used.
Since then it's just been a cycle of the old model being progressively lobotomised and a "new" one coming out that if you're lucky might be as good as the OG Opus 4.5 for a couple of weeks.
Subjective but as far as I can tell no progress in almost a year, which is a lifetime in 2022-25 LLM timelines
For comparison on openrouter DeepSeek v4 Flash is slightly cheaper than Gemma 4 31b, more expensive than Gemma 4 26b, but it does support prompt caching, which means for some applications it will be the cheapest. Excited to see how it compares with Gemma 4.
For those who rely on open source models but don't want to stop using frontier models, how do you manage it? Do you pay any of the Chinese subscription plans? Do you pay the API directly? After GPT 5.5 release, however good it is, I am a bit tired of this price hiking and reduced quota every week. I am now unemployed and cannot afford more expensive plans for the moment.
So, this is the version that's able to serve inference from Huawei chips, although it was still trained on nVidia. So unless I'm very much mistaken this is the biggest and best model yet served on (sort of) readily-available chinese-native tech. Performance and stability will be interesting to see; openrouter currently saying about 1.12s and 30tps, which isn't wonderful but it's day one after all.
For reference, the huawei Ascend 950 that this thing runs on is supposed to be roughly comparable to nVidia's H100 from 2022. In other words, things are hotting up in the GPU war!
I don't think we need to compare models to Opus anymore. Opus users don't care about other models, as they're convinced Opus will be better forever. And non-Opus users don't want the expense, lock-in or limits.
As a non-Opus user, I'll continue to use the cheapest fastest models that get my job done, which (for me anyway) is still MiniMax M2.5. I occasionally try a newer, more expensive model, and I get the same results. I have a feeling we might all be getting swindled by the whole AI industry with benchmarks that just make it look like everything's improving.
Which model's best depends on how you use it. There's a huge difference in behaviour between Claude and GPT and other models which makes some poor substitutes for others in certain use cases. I think the GPT models are a bad substitute for Claude ones for tasks such as pair-programming (where you want to see the CoT and have immediate responses) and writing code that you actually want to read and edit yourself, as opposed to just letting GPT run in the background to produce working code that you won't inspect. Yes, GPT 5.4 is cheap and brilliant but very black-box and often very slow IME. GPT-5.4 still seems to behave the same as 5.1, which includes problems like: doesn't show useful thoughts, can think for half an hour, says "Preparing the patch now" then thinks for another 20 min, gives no impression of what it's doing, reads microscopic parts of source files and misses context, will do anything to pass the tests including patching libraries...
Agree with your assessment, I think after models reached around Opus 4.5 level, its been almost indistinguishable for most tasks. Intelligence has been commoditized, what's important now is the workflows, prompting, and context management. And that is unique to each model.
Same for me. There are tasks when I want the smartest model. But for a whole lot of tasks I now default to Sonnet, or go with cheaper models like GLM, Kimi, Qwen. DeepSeek hasn't been in the mix for a while because their previous model had started lagging, but will definitely test this one again.
The tricky part is that the "number of tokens to good result" does absolutely vary, and you need a decent harness to make it work without too much manual intervention, so figuring out which model is most cost-effective for which tasks is becoming increasingly hard, but several are cost-effective enough.
Is Opus nerfed somehow in Copilot? Ive tried it numerous times, it has never reallt woved me. They seem to have awfully small context windows, but still. Its mostly their reasoning which has been off
Codex is just so much better, or the genera GPT models.
Try Charm Crush first, it's a native binary. If it's unbearable, try opencode, just with the knowledge your system will probably be pwned soon since it's JS + NPM + vibe coding + some of the most insufferable devs in the industry behind that product.
If you're feeling frisky, Zed has a decent agent harness and a very good editor.
actually this is not the reason - the harness is significantly better.
There is no comparable harness to Claude Code with skills, etc.
Opencode was getting there, but it seems the founders lost interest. Pi could be it, but its very focused on OpenClaw. Even Codex cli doesnt have all of it.
What's the issue with OC? I tried it a bit over 2 months ago, when I was still on Claude API, and it actually liked more that CC (i.e. the right sidebar with the plan and a tendency at asking less "security" questions that CC). Why is it so bad nowadays?
eh idk. until yesterday opus was the one that got spatial reasoning right (had to do some head pose stuff, neither glm 5.1 nor codex 5.3 could "get" it) and codex 5.3 was my champion at making UX work.
So while I agree mixed model is the way to go, opus is still my workhorse.
How does it compare to Opus 4.7? I've been immersed in 4.7 all week participating in the Anthropic Opus 4.7 hackathon and it's pretty impressive even if it's ravenous from a token perspective compared to 4.6
In theory, sure, but as other have pointed out you need to spend half a million on GPUs just to get enough VRAM to fit a single instance of the model. And you’d better make sure your use case makes full 24/7 use of all that rapidly-depreciating hardware you just spent all your money on, otherwise your actual cost per token will be much higher than you think.
In practice you will get better value from just buying tokens from a third party whose business is hosting open weight models as efficiently as possible and who make full use of their hardware. Even with the small margin they charge on top you will still come out ahead.
There are a lot of companies who would gladly drop half a million on a GPU to have private inference that Anthropic or OpenAI can’t use to steal their data.
And that GPU wouldn’t run one instance, the models are highly parallelizable. It would likely support 10-15 users at once, if a company oversubscribed 10:1 that GPU supports ~100 seats. Amortized over a couple years the costs are competitive.
> There are a lot of companies who would gladly drop half a million on a GPU to have private inference that Anthropic or OpenAI can’t use to steal their data.
Obviously, and certainly companies do run their own models because they place some value on data sovereignty for regulatory or compliance or other reasons. (Although the framing that Anthropic or OpenAI might "steal their data" is a bit alarmist - plenty of companies, including some with _highly_ sensitive data, have contracts with Anthropic or OpenAI that say they can't train future models on the data they send them and are perfectly happy to send data to Claude. You may think they're stupid to do that, but that's just your opinion.)
> the models are highly parallelizable. It would likely support 10-15 users at once.
Yes, I know that; I understand LLM internals pretty well. One instance of the model in the sense of one set of weights loaded across X number of GPUs; of course you can then run batch inference on those weights, up to the limits of GPU bandwidth and compute.
But are those 100 users you have on your own GPUs usings the GPUs evenly across the 24 hours of the day, or are they only using them during 9-5 in some timezone? If so, you're leaving your expensive hardware idle for 2/3 of the day and the third party providers hosting open weight models will still beat you on costs, even without getting into other factors like they bought their GPUs cheaper than you did. Do the math if you don't believe me.
To me, the important thing isn't that I can run it, it's that I can pay someone else to run it. I'm finding Opus 4.7 seems to be weirdly broken compared to 4.6, it just doesn't understand my code, breaks it whenever I ask it to do anything.
Now, at the moment, i can still use 4.6 but eventually Anthropic are going to remove it, and when it's gone it will be gone forever. I'm planning on trying Deepseek v4, because even if it's not quite as good, I know that it will be available forever, I'll always be able to find someone to run it.
No, but businesses do. Being able to run quality LLMs without your business, or business's private information, being held at the mercy of another corp has a lot of value.
But can be, and is, done. I work for a bootstrapped startup that hosts a DeepSeek v3 retrain on our own GPUs. We are highly profitable. We're certainly not the only ones in the space, as I'm personally aware of several other startups hosting their own GLM or DeepSeek models.
Completely agree, not suggesting it needs ot just genuinely curious. Love that it can be run locally though. Open source LLMs punching back pretty hard against proprietary ones in the cloud lately in terms of performance.
- To run with "heavy quantization" (16 bits -> 8): "8xH100", giving us $200K upfront and $4/h.
- To run truly "locally"--i.e. in a house instead of a data center--you'd need four 4090s, one of the most powerful consumer GPUs available. Even that would clock in around $15k for the cards alone and ~$0.22/h for the electricity (in the US).
Truly an insane industry. This is a good reminder of why datacenter capex from since 2023 has eclipsed the Manhattan Project, the Apollo program, and the US interstate system combined...
I remember reading about some new frameworks have been coming out to allow Macs to stream weights of huge models live from fast SSDs and produce quality output, albeit slowly. Apart from that...good luck finding that much available VRAM haha
It is more than good enough and has effectively caught up with Opus 4.6 and GPT 5.4 according to the benchmarks.
It's about 2 months behind GPT 5.5 and Opus 4.7.
As long as it is cheap to run for the hosting providers and it is frontier level, it is a very competitive model and impressive against the others. I give it 2 years maximum for consumer hardware to run models that are 500B - 800B quantized on their machines.
It should be obvious now why Anthropic really doesn't want you to run local models on your machine.
Vibes > Benchmarks. And it's all so task-specific. Gemini 3 has scored very well in benchmarks for very long but is poor at agentic usecases. A lot of people prefering Opus 4.6 to 4.7 for coding despite benchmarks, much more than I've seen before (4.5->4.6, 4->4.5).
Doesn't mean Deepseek v4 isn't great, just benchmarks alone aren't enough to tell.
Apparently glm5.1 and qwen coder latest is as good as opus 4.6 on benchmarks. So I tried both seriously for a week (glm Pro using CC) and qwen using qwen companion. Thought I could save $80 a month. Unfortunately after 2 days I had switched back to Max. The speed (slower on both although qwen is much faster) and errors (stupid layout mistakes, inserting 2 footers then refusing to remove one, not seeing obvious problems in screenshots & major f-ups of functionality), not being able to view URLs properly, etc. I'll give deepseek a go but I suspect it will be similar. The model is only half the story. Also been testing gpt5.4 with codex and it is very almost as good as CC... better on long running tasks running in background. Not keen on ChatGPT codex 'personality' so will stick to CC for the most part.
Their Chinese announcement says that, based on internal employee testing, it is not as good as Opus 4.6 Thinking, but is slightly better than Opus 4.6 without Thinking enabled.
That's super interesting, isn't Deepseek in China banned from using Anthropic models? Yet here they're comparing it in terms of internal employee testing.
They use VPN to access. Even Google Deepmind uses Anthropic. There was a fight within Google as to why only DeepMind is allowed to Claude while rest of the Google can't.
For the curious, I did some napkin math on their posted benchmarks and it racks up 20.1 percentage point difference across the 20 metrics where both were scored, for an average improvement of about 2% (non-pp). I really can't decide if that's mind blowing or boring?
Claude4.6 was almost 10pp better at at answering questions from long contexts ("corpuses" in CorpusQA and "multiround conversations" in MRCR), while DSv4 was a staggering 14pp better at one math challenge (IMOAnswerBench) and 12pp better at basic Q&A (SimpleQA-Verified).
There we go again :) It seems we have a release each day claiming that. What's weird is that even deepseek doesn't claim it's better than opus w/ thinking. No idea why you'd say that but anyway.
Dsv3 was a good model. Not benchmaxxed at all, it was pretty stable where it was. Did well on tasks that were ood for benchmarks, even if it was behind SotA.
This seems to be similar. Behind SotA, but not by much, and at a much lower price. The big one is being served (by ds themselves now, more providers will come and we'll see the median price) at 1.74$ in / 3.48$ out / 0.14$ cache. Really cheap for what it offers.
The small one is at 0.14$ in / 0.28$ out / 0.028$ cache, which is pretty much "too cheap to matter". This will be what people can run realistically "at home", and should be a contender for things like haiku/gemini-flash, if it can deliver at those levels.
> According to evaluation feedback, its user experience is better than Sonnet 4.5, and its delivery quality is close to Opus 4.6's non-thinking mode, but there is still a certain gap compared to Opus 4.6's thinking mode.
It is great! I asked the question what I always ask of new models ("what would Ian M Banks think about the current state of AI") and it gave me a brilliant answer! Funny enough the answer contained multiple criticisms of his own creators ("Chinese state entities", "Social Credit System").
Just tested it via openrounter in the Pi Coding agent and it regularly fails to use the read and write tool correctly, very disappointing. Anyone know a fix besides prompting "always use the provided tools instead of writing your own call"
The Flash version is 284B A13B in mixed FP8 / FP4 and the full native precision weights total approximately 154 GB. KV cache is said to take 10% as much space as V3. This looks very accessible for people running "large" local models. It's a nice follow up to the Gemma 4 and Qwen3.5 small local models.
You can, but does it work well? I assume CC has all kinds of Claude specific prompts in it, wouldn't you be better with a harness designed to be model agnostic like pi.dev or OpenCode?
This is just a random thought, but have you tried doing an 'agentic' pelican?
As in have the model consider its generated SVG, and gradually refine it, using its knowledge of the relative positions and proportions of the shapes generated, and have it spin for a while, and hopefully the end result will be better than just oneshotting it.
Or maybe going even one step further - most modern models have tool use and image recognition capabilities - what if you have it generate an SVG (or parts/layers of it, as per the model's discretion) and feed it back to itself via image recognition, and then improve on the result.
I think it'd be interesting to see, as for a lot of models, their oneshot capability in coding is not necessarily corellated with their in-harness ability, the latter which really matters.
I tried that for the GPT-5 launch - a self-improving loop that renders the SVG, looks at it and tries again - and the results were surprisingly disappointing.
I should try it again with the more recent models.
Being a bicycle geometry nerd I always look at the bicycle first.
Let me tell you how much the Pro one sucks... It looks like failed Pedersen[1]. The rear wheel intersects with the bottom bracket, so it wouldn't even roll. Or rather, this bike couldn't exist.
The flash one looks surprisingly correct with some wild fork offset and the slackest of seat tubes. It's got some lowrider[2] aspirations with the small wheels, but with longer, Rivendellish[3], chainstays. The seat post has different angle than the seat tube, so good luck lowering that.
This is an excellent comment. Thanks for this - I've only ever thought about whether the frame is the right shape, I never thought about how different illustrations might map to different bicycle categories.
I wonder which model will try some more common spoke lacing patterns. Right now there seems to be a preference for radial lacing, which is not super common (but simple to draw). The Flash and Pro one uses 16 spoke rims, which actually exist[1] but are not super common.
The Pro model fails badly at the spokes. Heck, the spokes sit on the outside of the drive side of the rim and tire. Have a nice ride riding on the spokes (instead of the tire) welded to the side of your rim.
Both bikes have the drive side on the left, which is very very uncommon. That can't exist in the training data.
I think the pelican on a bike is known widely enough that of seizes to be useful as a benchmark. There is even a pelican briefly appearing in the promo video of GPT-5, if I'm not mistaken https://openai.com/gpt-5/. So the companies are apparently aware of it.
Feels like the real story here is cost/performance tradeoff rather than raw capability. Benchmarks keep moving incrementally, but efficiency gains like this actually change who can afford to build on top.
I don't mind that High Flyer completely ripped off Anthropic to do this so much as I mind that they very obviously waited long enough for the GAB to add several dozen xz-level easter eggs to it.
Funny how Gemini is theoretically the best -- but in practice all the bugs in the interface mean I don't want to use it anymore. The worst is it forgets context (and lies about it), but it's very unreliable at reading pdfs (and lies about it). There's also no branch, so once the context is lost/polluted, you have to start projects over and build up the context from scratch again.
This is shockingly cheap for a near frontier model. This is insane.
For context, for an agent we're working on, we're using 5-mini, which is $2/1m tokens. This is $0.30/1m tokens. And it's Opus 4.6 level - this can't be real.
I am uncomfortable about sending user data which may contain PII to their servers in China so I won't be using this as appealing as it sounds. I need this to come to a US-hosted environment at an equivalent price.
Hosting this on my own + renting GPUs is much more expensive than DeepSeek's quoted price, so not an option.
At this point 'frontier model release' is a monthly cadence, Kimi 2.6 Claude 4.6 GPT 5.5, the interesting question is which evals will still be meaningful in 6 months.
SOTA MRCR (or would've been a few hours earlier... beaten by 5.5), I've long thought of this as the most important non-agentic benchmark, so this is especially impressive. Beats Opus 4.7 here
"Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 supernodes are launched at scale in the second half of this year, the price of Pro is expected to be reduced significantly."
For flash? 4 bit quant, 2x 96GB gpu (fast and expensive) or 1x 96GB gpu + 128GB ram (still expensive but probably usable, if you’re patient).
A mac with 256 GB memory would run it but be very slow, and so would be a 256GB ram + cheapo GPU desktop, unless you leave it running overnight.
The big model? Forget it, not this decade. You can theoretically load from SSD but waiting for the reply will be a religious experience.
Realistically the biggest models you can run on local-as-in-worth-buying-as-a-person hardware are between 120B and 200B, depending on how far you’re willing to go on quantization. Even this is fairly expensive, and that’s before RAM went to the moon.
There is no BF16. There is no FP8 for the instruct model. The instruct model at full precision is 160 GB (mixed FP4 and FP8). The base model at full precision is 284 GB (FP8). Almost everyone is going to use instruct. But I do love to see base models released.
Run on an old HEDT platform with a lot of parallel attached storage (probably PCIe 4) and fetch weights from SSD. You'd ultimately be limited by the latency of these per-layer fetches, since MoE weights are small. You could reduce the latencies further by buying cheap Optane memory on the second-hand market.
The low end could be something like an eBay-sourced server with a truckload of DDR3 ram doing all-cpu inference - secondhand server models with a terabyte of ram can be had for about 1.5K. The TPS will be absolute garbage and it will sound like a jet engine, but it will nominally run.
The flash version here is 284B A13B, so it might perform OK with a fairly small amount of VRAM for the active params and all regular ram for the other params, but I’d have to see benchmarks. If it turns out that works alright, an eBay server plus a 3090 might be the bang-for-buck champ for about $2.5K (assuming you’re starting from zero).
Was expecting that the release would be this month [1], since everyone forgot about it and not reading the papers they were releasing and 7 days later here we have it.
One of the key points of this model to look at is the optimization that DeepSeek made with the residual design of the neural network architecture of the LLM, which is manifold-constrained hyper-connections (mHC) which is from this paper [2], which makes this possible to efficiently train it, especially with its hybrid attention mechanism designed for this.
There was not that much discussion around it some months ago here [3] about it but again this is a recommended read of the paper.
I wouldn't trust the benchmarks directly, but would wait for others to try it for themselves to see if it matches the performance of frontier models.
Either way, this is why Anthropic wants to ban open weight models and I cannot wait for the quantized versions to release momentarily.
More like he wants to ban accelerator chip sales to China, which may be about “national security” or self preservation against a different model for AI development which also happens to be an existential threat to Anthropic. Maybe those alternatives are actually one and the same to him.
Using it with opencode sometimes it generates commands like:
bash({"command":"gh pr create --title "Improve Calendar module docs and clean up idiomatic Elixir" --body "$(cat <<'EOF'
Problem
The Calendar modu...
like generating output, but not actually running the bash command so not creating the PR ultimately. I wonder if it's a model thing, or an opencode thing.
How long does it usually take for folks to make smaller distills of these models? I really want to see how this will do when brought down to a size that will run on a Macbook.
Weren't there some frameworks recently released to allow Macs to stream weights from fast SSDs and thus fit way more parameters than what would normally fit in RAM?
I have never tried one yet but I am considering trying that for a medium sized model.
I've been calling that the "streaming experts" trick, the key idea is to take advantage of Mixture of Expert models where only a subset of the weights are used for each round of calculations, then load those weights from SSD into RAM for each round.
As I understand it if DeepSeek v4 Pro is a 1.6T, 49B active that means you'd need just 49B in memory, so ~100GB at 16 bit or ~50GB at 8bit quantized.
v4 Flash is 284B, 13B active so might even fit in <32GB.
The "active" count is not very meaningful except as a broad measure of sparsity, since the experts in MoE models are chosen per layer. Once you're streaming experts from disk, there's nothing that inherently requires having 49B parameters in memory at once. Of course, the less caching memory does, the higher the performance overhead of fetching from disk.
Streaming weights from RAM to GPU for prefill makes sense due to batching and pcie5 x16 is fast enough to make it worthwhile.
Streaming weights from RAM to GPU for decode makes no sense at all because batching requires multiple parallel streams.
Streaming weights from SSD _never_ makes sense because the delta between SSD and RAM is too large. There is no situation where you would not be able to fit a model in RAM and also have useful speeds from SSD.
These are more like experiments than a polished release as of yet. And the reduction in throughput is high compared to having the weights in RAM at all times, since you're bottlenecked by the SSD which even at its fastest is much slower than RAM.
But if it does, then in the following week we'll see DeepSeek4 floods every AI-related online space. Thousands of posts swearing how it's better than the latest models OpenAI/Anthropic/Google have but only costs pennies.
Then a few weeks later it'll be forgotten by most.
It's difficult because even if the underlying model is very good, not having a pre-built harness like Claude Code makes it very un-sticky for most devs. Even at equal quality, the friction (or at least perceived friction) is higher than the mainstream models.
If one finds it difficult to set up OpenCode to use whatever providers they want, I won't call them 'dev'.
The only real friction (if the model is actually as good as SOTA) is to convince your employer to pay for it. But again if it really provides the same value at a fraction of the cost, it'll eventually cease to be an issue.
"If one finds it difficult to set up OpenCode to use whatever providers they want, I won't call them 'dev'."
I feel the same way. But look at the ollama vs llama.cpp post from HN few days back and you will see most of the enthusiasts in this space are very non technical people.
Also, note that there's zero CUDA dependency. It runs entirely on Huawei chips. In other words, Chinese ecosystem has delivered a complete AI stack. Like it or not, that's a big news. But what's there not to like when monopolies break down?
China is not perfect but a bit of competition is healthy and needed
https://api-docs.deepseek.com/guides/thinking_mode
No BS, just a concise description of exactly what I need to write my own agent.
Western Models are optimizing to be used as an interchangeable product. Chinese models are being optimizing to be built upon.
Why? It sounds like the stupidest idea ever. Interchangeability = no lock-in = no moot.
They are developing their moats with the platform tooling around it right now though. Look at Anthropic with Routines and OpenAI with Agents. Drop that capability in to a business with loose controls and suddenly you have a very sticky product with high switching costs. Meanwhile if you stick with purely the ‘chat’ use cases, even Cowork and scheduled tasks, you maintain portability.
I have a collection of novel probability and statistics problems at the masters and PhD level with varying degrees of feasibility. My test suite involves running these problems through first (often with about 2-6 papers for context) and then requesting a rigorous proof as followup. Since the problems are pretty tough, there is no quantitative measure of performance here, I'm just judging based on how useful the output is toward outlining a solution that would hopefully become publishable.
Just prior to this model, Gemini led the pack, with GPT-5 as a close second. No other model came anywhere near these two (no, not even Claude). Gemini would sometimes have incredible insight for some of the harder problems (insightful guesses on relevant procedures are often most useful in research), but both of them tend to struggle with outlining a concrete proof in a single followup prompt. This DeepSeek V4 Pro with max thinking does remarkably well here. I'm not seeing the same level of insights in the first response as Gemini (closer to GPT-5), but it often gets much better in the followup, and the proofs can be _very_ impressive; nearly complete in several cases.
Given that both Gemini and DeepSeek also seem to lead on token performance, I'm guessing that might play a role in their capacity for these types of problems. It's probably more a matter of just how far they can get in a sensible computational budget.
Despite what the benchmarks seem to show, this feels like a huge step up for open-weight models. Bravo to the DeepSeek team!
Pretty cool, I think they're the first to guarantee determinism with the fixed seed or at the temperature 0. Google came close but never guaranteed it AFAIK. DeepSeek show their roots - it may not strictly be a SotA model, but there's a ton of low-level optimizations nobody else pays attention to.
I’d like somebody to explain to me how the endless comments of "bleeding edge labs are subsidizing the inference at an insane rate" make sense in light of a humongous model like v4 pro being $4 per 1M. I’d bet even the subscriptions are profitable, much less the API prices.
edit: $1.74/M input $3.48/M output on OpenRouter
But seriously, it just stems from the fact some people want AI to go away. If you set your conclusion first, you can very easily derive any premise. AI must go away -> AI must be a bad business -> AI must be losing money.
Aka: everyone who uses Nvidia isn't selling at cost, because Nvidia is so expensive.
One answer - Chinese Communist Party. They are being subsidized by the state.
Back in Nov 2025, Opus 4.5 (80.9%) was the first proprietary model to do so.
So it os hard to tell how much of a model gain is due to skill, and how much - overfitting.
Edit: it seems "open source" was edited out of the parent comment.
no one is ever going to release their training data because it contains every copyrighted work in existence. everyone, even the hecking-wholesome safety-first Anthropic, is using copyrighted data without permission to train their models. there you go.
It is very much a valuable thing already, no need to taint it with wrong promise.
Though I disagree about being used if it was indeed open source: I might not do it inside my home lab today, but at least Qwen and DeepSeek would use and build on what eg. Facebook was doing with Llama, and they might be pushing the open weights model frontier forward faster.
https://www.reuters.com/technology/nvidia-is-sued-by-authors...
1. Training data is the source. 2. Training is compilation/compression. 3. Weights are the compiled source akin to optimized assembly.
However it's an imperfect analogy on so many levels. Nitpick away.
[0] https://news.ycombinator.com/item?id=47758408
This “no harm to me” meme about a foreign totalitarian government (with plenty of incentive to run influence ops on foreigners) hoovering your data is just so mind-bogglingly naive.
Relatively speaking, DeepSeek is less untrustworthy than Grok.
When I try ChatGPT on current events from the White House it interprets them as strange hypotheticals rather than news, which is probably more a problem with DC than with GPT, but whatever.
Even for minor stuff like beeing addicted to drugs.
Looks pretty totalitarian to me.
Note: you can have this conversation criticizing the US on a US website. Try criticizing Xi or the CCP or calling him Pooh on a Chinese website.
You think China doesn’t imprison drug users?
China recently executed a low level drug trafficker
https://www.lemonde.fr/en/international/article/2026/04/05/c...
China is one of the top executioners. China executes more than rest of the world combined
https://www.amnesty.org/en/latest/news/2017/04/china-must-co...
You think China is honest about political prisoners in Tibet and Xinjiang?
Criticize the US all you want but I can’t understand the whitewashing of a real totalitarian and genocidal state like mainland China.
That would be a great argument if the American models weren’t so heavily censored.
The Chinese model might dodge a question if I ask it about 1-2 specific Chinese cultural issues but then it also doesn’t moralize me at every turn because I asked it to use a piece of security software.
yes, this is exactly what I'm saying.
This is why I’ve been urging everyone I know to move away from American based services and providers. It’s slow but honest work.
The executive branch?
My country’s per capita income is $2500 a year. We can’t pay perpetual rent to OAI/Anthropic
This sounds whole lot like potatoh potahto. I think the former argument is very much the correct one: China can undercut everyone and win, even at a loss. Happened with solar panels, steel, evs, sea food - it's a well tested strategy and it works really well despite the many flavors it comes in.
That being said a job well done for the wrong reasons is still a job well done so we should very much welcome these contributions, and maybe it's good to upset western big tech a bit so it's remains competitive.
Just this week they published a serious foundational library for LLMs https://github.com/deepseek-ai/TileKernels
Others worth mentioning:
https://github.com/deepseek-ai/DeepGEMM a competitive foundational library
https://github.com/deepseek-ai/Engram
https://github.com/deepseek-ai/DeepSeek-V3
https://github.com/deepseek-ai/DeepSeek-R1
https://github.com/deepseek-ai/DeepSeek-OCR-2
They have 33 repos and counting: https://github.com/orgs/deepseek-ai/repositories?type=all
And DeepSeek often has very cool new approaches to AI copied by the rest. Many others copied their tech. And some of those have 10x or 100x the GPU training budget and that's their moat to stay competitive.
The models from Chinese Big Tech and some of the small ones are open weights only. (and allegedly benchmaxxed) (see https://xcancel.com/N8Programs/status/2044408755790508113). Not the same.
So you can’t see what facts are pruned out, what biases were applied, etc. Even more importantly, you can’t make a slightly improved version.
This model is as open source as a windows XP installation ISO.
https://openrouter.ai/deepseek/deepseek-v4-flash
`https://openrouter.ai/api/messages with model=deepseek/deepseek-v4-pro, OR returns an error because their Anthropic-compat translator doesn't cover V4 yet. The Claude CLI dutifully surfaces that error as "model...does not exist"
At this point I would just pick the one who's "ethics" and user experience you prefer. The difference in performance between these releases has had no impact on the meaningful work one can do with them, unless perhaps they are on the fringes in some domain.
Personally I am trying out the open models cloud hosted, since I am not interested in being rug pulled by the big two providers. They have come a long way, and for all the work I actually trust to an LLM they seem to be sufficient.
New model comes out, has some nice benchmarks, but the subjective experience of actually using it stays the same. Nothing's really blown my mind since.
Feels like the field has stagnated to a point where only the enthusiasts care.
Since then it's just been a cycle of the old model being progressively lobotomised and a "new" one coming out that if you're lucky might be as good as the OG Opus 4.5 for a couple of weeks.
Subjective but as far as I can tell no progress in almost a year, which is a lifetime in 2022-25 LLM timelines
For reference, the huawei Ascend 950 that this thing runs on is supposed to be roughly comparable to nVidia's H100 from 2022. In other words, things are hotting up in the GPU war!
And we got new base models, wonderful, truly wonderful
Model was released and it's amazing. Frontier level (better than Opus 4.6) at a fraction of the cost.
As a non-Opus user, I'll continue to use the cheapest fastest models that get my job done, which (for me anyway) is still MiniMax M2.5. I occasionally try a newer, more expensive model, and I get the same results. I have a feeling we might all be getting swindled by the whole AI industry with benchmarks that just make it look like everything's improving.
The tricky part is that the "number of tokens to good result" does absolutely vary, and you need a decent harness to make it work without too much manual intervention, so figuring out which model is most cost-effective for which tasks is becoming increasingly hard, but several are cost-effective enough.
Substantially worse at following instructions and overoptimized for maximizing token usage
Codex is just so much better, or the genera GPT models.
https://github.blog/news-insights/company-news/changes-to-gi...
I do some stuff with gemini flash and Aider, but mostly because I want to avoid locking myself into a walled garden of models, UIs and company
If you're feeling frisky, Zed has a decent agent harness and a very good editor.
Opencode was getting there, but it seems the founders lost interest. Pi could be it, but its very focused on OpenClaw. Even Codex cli doesnt have all of it.
which harness works well with Deepseek v4 ?
So while I agree mixed model is the way to go, opus is still my workhorse.
> In our internal evaluation, DeepSeek-V4-Pro-Max outperforms Claude Sonnet 4.5 and approaches the level of Opus 4.5.
This is free... as in you can download it, run it on your systems and finetune it to be the way you want it to be.
In theory, sure, but as other have pointed out you need to spend half a million on GPUs just to get enough VRAM to fit a single instance of the model. And you’d better make sure your use case makes full 24/7 use of all that rapidly-depreciating hardware you just spent all your money on, otherwise your actual cost per token will be much higher than you think.
In practice you will get better value from just buying tokens from a third party whose business is hosting open weight models as efficiently as possible and who make full use of their hardware. Even with the small margin they charge on top you will still come out ahead.
And that GPU wouldn’t run one instance, the models are highly parallelizable. It would likely support 10-15 users at once, if a company oversubscribed 10:1 that GPU supports ~100 seats. Amortized over a couple years the costs are competitive.
Obviously, and certainly companies do run their own models because they place some value on data sovereignty for regulatory or compliance or other reasons. (Although the framing that Anthropic or OpenAI might "steal their data" is a bit alarmist - plenty of companies, including some with _highly_ sensitive data, have contracts with Anthropic or OpenAI that say they can't train future models on the data they send them and are perfectly happy to send data to Claude. You may think they're stupid to do that, but that's just your opinion.)
> the models are highly parallelizable. It would likely support 10-15 users at once.
Yes, I know that; I understand LLM internals pretty well. One instance of the model in the sense of one set of weights loaded across X number of GPUs; of course you can then run batch inference on those weights, up to the limits of GPU bandwidth and compute.
But are those 100 users you have on your own GPUs usings the GPUs evenly across the 24 hours of the day, or are they only using them during 9-5 in some timezone? If so, you're leaving your expensive hardware idle for 2/3 of the day and the third party providers hosting open weight models will still beat you on costs, even without getting into other factors like they bought their GPUs cheaper than you did. Do the math if you don't believe me.
Now, at the moment, i can still use 4.6 but eventually Anthropic are going to remove it, and when it's gone it will be gone forever. I'm planning on trying Deepseek v4, because even if it's not quite as good, I know that it will be available forever, I'll always be able to find someone to run it.
If you want to go budget corporate, 7 x H200 is just barely going to run it, but all in, $300k ought to do it.
- To run at full precision: "16–24 H100s", giving us ~$400-600k upfront, or $8-12/h from [us-east-1](https://intuitionlabs.ai/articles/h100-rental-prices-cloud-c...).
- To run with "heavy quantization" (16 bits -> 8): "8xH100", giving us $200K upfront and $4/h.
- To run truly "locally"--i.e. in a house instead of a data center--you'd need four 4090s, one of the most powerful consumer GPUs available. Even that would clock in around $15k for the cards alone and ~$0.22/h for the electricity (in the US).
Truly an insane industry. This is a good reminder of why datacenter capex from since 2023 has eclipsed the Manhattan Project, the Apollo program, and the US interstate system combined...
10 years from now that hardware will be on eBay for any geek with a couple thousand dollars and enough power to run it.
"671B total / 37B active"
"Full precision (BF16)"
And they claim they ran this non-existent model on vLLM and SGLang over a month and a half ago.
It's clickbait keyword slop filled in with V3 specs. Most of the web is slop like this now. Sigh.
It's about 2 months behind GPT 5.5 and Opus 4.7.
As long as it is cheap to run for the hosting providers and it is frontier level, it is a very competitive model and impressive against the others. I give it 2 years maximum for consumer hardware to run models that are 500B - 800B quantized on their machines.
It should be obvious now why Anthropic really doesn't want you to run local models on your machine.
Doesn't mean Deepseek v4 isn't great, just benchmarks alone aren't enough to tell.
If its coding abilities are better than Claude Code with Opus 4.6 then I will definitely be switching to this model.
It's still a "preview" version atm.
Claude4.6 was almost 10pp better at at answering questions from long contexts ("corpuses" in CorpusQA and "multiround conversations" in MRCR), while DSv4 was a staggering 14pp better at one math challenge (IMOAnswerBench) and 12pp better at basic Q&A (SimpleQA-Verified).
There we go again :) It seems we have a release each day claiming that. What's weird is that even deepseek doesn't claim it's better than opus w/ thinking. No idea why you'd say that but anyway.
Dsv3 was a good model. Not benchmaxxed at all, it was pretty stable where it was. Did well on tasks that were ood for benchmarks, even if it was behind SotA.
This seems to be similar. Behind SotA, but not by much, and at a much lower price. The big one is being served (by ds themselves now, more providers will come and we'll see the median price) at 1.74$ in / 3.48$ out / 0.14$ cache. Really cheap for what it offers.
The small one is at 0.14$ in / 0.28$ out / 0.028$ cache, which is pretty much "too cheap to matter". This will be what people can run realistically "at home", and should be a contender for things like haiku/gemini-flash, if it can deliver at those levels.
LMAO
I have no idea why you'd think that, but this is straight from their announcement here (https://mp.weixin.qq.com/s/8bxXqS2R8Fx5-1TLDBiEDg):
> According to evaluation feedback, its user experience is better than Sonnet 4.5, and its delivery quality is close to Opus 4.6's non-thinking mode, but there is still a certain gap compared to Opus 4.6's thinking mode.
This is the model creators saying it, not me.
That's literally what the I Ching calls "good fortune."
Competition, when no single dragon monopolizes the sky, brings fortune for all.
The website now has a link to the announcement on Twitter here https://x.com/deepseek_ai/status/2047516922263285776
Copying text of that below
DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.
DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.
Try it now at http://chat.deepseek.com via Expert Mode / Instant Mode. API is updated & available today!
Tech Report: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main...
Open Weights: https://huggingface.co/collections/deepseek-ai/deepseek-v4
https://api-docs.deepseek.com/guides/coding_agents#integrate...
But in this case, it's more likely just to be a tooling issue.
input: $0.14/$0.28 (whereas gemini $0.5/$3)
Does anyone know why output prices have such a big gap?
https://simonwillison.net/2026/Apr/24/deepseek-v4/
Both generated using OpenRouter.
For comparison, here's what I got from DeepSeek 3.2 back in December: https://simonwillison.net/2025/Dec/1/deepseek-v32/
And DeepSeek 3.1 in August: https://simonwillison.net/2025/Aug/22/deepseek-31/
And DeepSeek v3-0324 in March last year: https://simonwillison.net/2025/Mar/24/deepseek/
As in have the model consider its generated SVG, and gradually refine it, using its knowledge of the relative positions and proportions of the shapes generated, and have it spin for a while, and hopefully the end result will be better than just oneshotting it.
Or maybe going even one step further - most modern models have tool use and image recognition capabilities - what if you have it generate an SVG (or parts/layers of it, as per the model's discretion) and feed it back to itself via image recognition, and then improve on the result.
I think it'd be interesting to see, as for a lot of models, their oneshot capability in coding is not necessarily corellated with their in-harness ability, the latter which really matters.
I should try it again with the more recent models.
Let me tell you how much the Pro one sucks... It looks like failed Pedersen[1]. The rear wheel intersects with the bottom bracket, so it wouldn't even roll. Or rather, this bike couldn't exist.
The flash one looks surprisingly correct with some wild fork offset and the slackest of seat tubes. It's got some lowrider[2] aspirations with the small wheels, but with longer, Rivendellish[3], chainstays. The seat post has different angle than the seat tube, so good luck lowering that.
[1] https://en.wikipedia.org/wiki/Pedersen_bicycle
[2] https://en.wikipedia.org/wiki/Lowrider_bicycle
[3] https://www.rivbike.com/
I wonder which model will try some more common spoke lacing patterns. Right now there seems to be a preference for radial lacing, which is not super common (but simple to draw). The Flash and Pro one uses 16 spoke rims, which actually exist[1] but are not super common.
The Pro model fails badly at the spokes. Heck, the spokes sit on the outside of the drive side of the rim and tire. Have a nice ride riding on the spokes (instead of the tire) welded to the side of your rim.
Both bikes have the drive side on the left, which is very very uncommon. That can't exist in the training data.
[1] https://cicli-berlinetta.com/product/campagnolo-shamal-16-sp...
1) LLM is not AGI. Because surely if AGI it would imply that pro would do better than flash?
2) and because of the above, Pelican example is most likely already being benchmaxxed.
How much does the drawing change if you ask it again?
at the top of the linked pages.
Gemini-3.1-Pro at 91.0
Opus-4.6 at 89.1
GPT-5.4, Kimi2.6, and DS-V4-Pro tied at 87.5
Pretty impressive
If AI was so good at coding, why can’t it actually make a usable Gemini/AI Studio app?
For context, for an agent we're working on, we're using 5-mini, which is $2/1m tokens. This is $0.30/1m tokens. And it's Opus 4.6 level - this can't be real.
I am uncomfortable about sending user data which may contain PII to their servers in China so I won't be using this as appealing as it sounds. I need this to come to a US-hosted environment at an equivalent price.
Hosting this on my own + renting GPUs is much more expensive than DeepSeek's quoted price, so not an option.
As a European I feel deeply uncomfortable about sending data to US companies where I know for sure that the government has access to it.
I also feel uncomfortable sending it to China.
If you'd asked me ten years ago which one made me more uncomfortable. China.
But now I'm not so sure, in fact I'm starting to lean towards the US as being the major risk.
dang, probably the two should be merged and that be the link
"Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 supernodes are launched at scale in the second half of this year, the price of Pro is expected to be reduced significantly."
So it's going to be even cheaper
A mac with 256 GB memory would run it but be very slow, and so would be a 256GB ram + cheapo GPU desktop, unless you leave it running overnight.
The big model? Forget it, not this decade. You can theoretically load from SSD but waiting for the reply will be a religious experience.
Realistically the biggest models you can run on local-as-in-worth-buying-as-a-person hardware are between 120B and 200B, depending on how far you’re willing to go on quantization. Even this is fairly expensive, and that’s before RAM went to the moon.
The flash version here is 284B A13B, so it might perform OK with a fairly small amount of VRAM for the active params and all regular ram for the other params, but I’d have to see benchmarks. If it turns out that works alright, an eBay server plus a 3090 might be the bang-for-buck champ for about $2.5K (assuming you’re starting from zero).
Was expecting that the release would be this month [1], since everyone forgot about it and not reading the papers they were releasing and 7 days later here we have it.
One of the key points of this model to look at is the optimization that DeepSeek made with the residual design of the neural network architecture of the LLM, which is manifold-constrained hyper-connections (mHC) which is from this paper [2], which makes this possible to efficiently train it, especially with its hybrid attention mechanism designed for this.
There was not that much discussion around it some months ago here [3] about it but again this is a recommended read of the paper.
I wouldn't trust the benchmarks directly, but would wait for others to try it for themselves to see if it matches the performance of frontier models.
Either way, this is why Anthropic wants to ban open weight models and I cannot wait for the quantized versions to release momentarily.
[0] https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main...
[1] https://news.ycombinator.com/item?id=47793880
[2] https://arxiv.org/abs/2512.24880
[3] https://news.ycombinator.com/item?id=46452172
Do you have a source?
Keep an eye on https://huggingface.co/unsloth/models
Update ten minutes later: https://huggingface.co/unsloth/DeepSeek-V4-Pro just appeared but doesn't have files in yet, so they are clearly awake and pushing updates.
I have never tried one yet but I am considering trying that for a medium sized model.
As I understand it if DeepSeek v4 Pro is a 1.6T, 49B active that means you'd need just 49B in memory, so ~100GB at 16 bit or ~50GB at 8bit quantized.
v4 Flash is 284B, 13B active so might even fit in <32GB.
My Mac can fit almost 70B (Q3_K_M) in memory at once, so I really need to try this out soon at maybe Q5-ish.
V4 is natively mixed FP4 and FP8, so significantly less than that. 50 GB max unquantized.
Streaming weights from RAM to GPU for decode makes no sense at all because batching requires multiple parallel streams.
Streaming weights from SSD _never_ makes sense because the delta between SSD and RAM is too large. There is no situation where you would not be able to fit a model in RAM and also have useful speeds from SSD.
Note: these were just two that I starred when I saw them posted here. I have not looked seriously at it at the moment,
https://github.com/danveloper/flash-moe
https://github.com/t8/hypura
https://news.ycombinator.com/item?id=47885014
https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
OMG ITS HAPPENING
But if it does, then in the following week we'll see DeepSeek4 floods every AI-related online space. Thousands of posts swearing how it's better than the latest models OpenAI/Anthropic/Google have but only costs pennies.
Then a few weeks later it'll be forgotten by most.
If one finds it difficult to set up OpenCode to use whatever providers they want, I won't call them 'dev'.
The only real friction (if the model is actually as good as SOTA) is to convince your employer to pay for it. But again if it really provides the same value at a fraction of the cost, it'll eventually cease to be an issue.
Damn autocorrect :)