DeepSeek V4 Flash 0731
arcprize.org
arcprize.org
OpenCode Go even has double limits temporarily so for 10 USD you effectively get 140 USD of tokens to spend. It would impress me if someone could burn that amount with "normal" usage. Even when running multiple sessions.
I have a Claude Max subscription but I've barely touched it, it just feels like a step back to have to think about limits and usage even though the models are stronger.
The beauty of intelligence at this cost (even if it's not SOTA) is that it opens a whole bunch of new use cases. Test failure in CI? Have the bot automatically propose a fix, its cheap enough that you can discard it w/h issues. Test coverage too low? Auto generate tests on CI for every pull-requests! Monitoring server logs, continuous security audits and investigating every received exception now becomes possible.
I'm thinking about having it automatically filter and re-rank my social media feeds so I can steer the algorithm instead of the other way around.
Perhaps other people (with enormous budgets) were already doing all of the above but for us this is a really exciting release!
"We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice."
vLLM has recently released a similar approach. It's not as effective as what DeepSeek does but still an interesting development.
I have no doubt that in due time other providers will match or perhaps even beat the current DeepSeek prices.
- input_cache_hit_tokens: 1,265,646,976 x 0.0000000028 = $3.5438115328
- input_cache_miss_tokens: 18,208,088 x 0.00000014 = $2.54913232
- output_tokens: 9,615,178 x 0.00000028 = $2.69224984
- request_count: 10,837 (no price)
Cache:
- Hit: 1,265,646,976
- Miss: 18,208,088
- Total input tokens: 1,283,855,064
Deepseek charges $0.0028 per cache read on Openrouter. The next cheapest is $0.018.
That's a massive difference and quickly adds up on coding sessions (which often hit 95%+ cached tokens).
This adds disk as a tier in the HBM → CPU → Disk KV cache hierarchy.
There's also a cluster of related KV-offload FS PRs: #49225 (read/write batching, still open) and #49152 (batch store/load in C, merged Jul 28).
It's hard to say if these are similar to the approach DeepSeek takes but they definitely seem very interesting.
The entire issue is caching, I tried to write some custom to dump to disk kv-caching using some ideas from their papers and my experience with snapshots and vm checkpoint systems, I must say they must have really squeezed that lemon it's hard.
Atleast me with Sol couldn't figure it out over a couple days, a few hours each day, which isn't much but I did feel a bit stuck with existing solutions and felt like I might have to write something from scratch. But if you are willing to put in the effort into the infra I do think it's doable. But it will be really hard to pull it off.
My congrats to anyone who manages to pull it off, they might be able to kill off most AI labs. Assuming they can find the compute, Deepseek really has killed all models for me other than Sol/Fable/Opus/K3 tier stuff.
And there is no way in hell anyone can afford caching prices same as what DeepSeek is offering, and DeepSeek keeps the cache available for an insane amount of time most providers will flush it in 5-mins like Claude/Anthropic (some offer customizing it but I am not sure of the pricing, it's load based on some like Fireworks, which means assume a couple minutes at most, they say several minutes god knows what that really means).
There is no way to match DeepSeek's current prices, "profitably" if you are renting a GPU and reselling tokens, unless you have some really amazing caching infra or something.
Deepseek's prices are just insanely cheap, I am not saying it's impossible to get there the overall performance suggests it should be feasible, but I will be damned if any provider could match their tps and caching any time soon at those same prices profitably.
I believe even if Deepseek 2-3x their prices across the board even then they would be cheaper for most long running tasks, that's just how good their caching is.
For one I have managed to hit the cache after over 24 hours on their system it's insane, I honestly didn't care because it was so cheap but it truly made me incredibly happy to think about the engineering that must have taken. TTFT is slightly worse, but it's good enough, for those cache prices I can take a few seconds worth of hit on TTFT.
Can anyone working at one of the main US labs (Google, OpenAI, Anthropic) comment on WTF they haven't even tried MLA - despite the obvious massive advantages?
I know enough to know they aren't completely incompetent. So there must be a quite good reason.
But it remains a mystery to me.
DeepSeek's MLA is like almost 2 years old at this time. They've got thousands of people working on this stuff. They clearly have the ability to at least try it...
The big US labs are opaque and don't publish much of any technical details anymore. We don't know what they are or aren't doing, honestly.
There’s a measurable performance tradeoff versus gqa so there’s reluctance.
For the most part though the new deepseek v4 tech is hca and mhc and people are still catching on like with moe and rl. Wait for 6 12 months, minimum time for next pre train.
It's interesting that most open models adding 1M context did it in a way that reduces KV cache size (though DeepSeek was the most aggressive, using compressed attention on all layers), but only a couple providers turned it into a discount on cache reads.
They "can" is the caveat here. Rented GPUs are going up in pricing. I recently got an email that DigitalOcean pricing of GPUs were going up.
So
1. They have to get a hold of them (availability is bad)
2. They have to maintain the pricing
[0] https://openrouter.ai/deepseek/deepseek-v4-flash-0731#provid...
And, probably 99.99% of people using LLM probably don't even need SOTA anyway.
I hadn't really thought about this but AI may well be the technology that disrupts and ultimately destroys social media.
The value proposition of something like FB or IG is, as we know, the network effect. The platform gets to extract value from user generated content. I believe that users should own the platform, a bit like the Wikimedia Foundation, because they're the ones that create value. Federation is a popular belief on HN and I've come to believe that's simply the wrong solution to the right problem.
Anyway, how these social media companies make money is by optimizing the feed for engagement. People know it too so you see people trying to build an audience by rage baiting. And then more time spent equals more advertising revenue.
But what happens when the AI can simply slurp all the posts and then filter and rank them? It destroys the engagement and advertising model. And I'm not opposed to that, honestly. It may be on eof the few good thing sto come out of AI.
Terrible use-case.
this seems like such a bad idea
"If this PR adds any new endpoints, ensure that there are functional and integration tests. If there are not, please investigate the feasibility and appropriateness, and create functional tests using the guide found on our wiki for guidance https://www.ourdevwiki.site/how-to-make-functional-tests" then maybe it could add some value.
But that very much depends on the specific system. Some tests are obvious, some not so much.
which in your case is?
My family uses it. I have gallery apps (yearbooks for each year are a lot of fun!) of us on trips and just living, an outlining app that's a mesh of Workflowy and Org Mode (it's called Fluxtral), a markdown-backed app (it uses marked.min.js, and is called Dextral) that offers documents, logs, calendars, and kanban boards, all parsed from markdown. I have a List app for gear, trips, shopping, etc. that we all can contribute to. There are utilities (world clock, calendar) and games (an oracle for RPGs, a KenKen implementation), and apps (a diagram editor that exports to SVG, a web-launcher that uses pneumonics, a Scheme-based hacking environment, and a spreadsheet that does most of what you'd expect aside from Solver and Pivot tables).
I started these projects before AI, and made slow progress over the years, but the modern versions of all this stuff have been built with Deepseek V4 Flash. I've also used Gemini in the very early days, and Kimi K2.6 later on, but these days, since I can now host Deepseek v4 Flash 0731 in a 2-bit quant on my Strix Halo box (128GB, but only about 250GB/s of memory bandwidth, so 15t/s), I used Deepseek with omp for almost everything. It's a very capable model, and I'm amazed I can run it locally and get good results. It's really revolutionary for my (small) use cases.
With unsloth's Q3_S quant + kyuz0 'llama-vulkan-radv-performance' toolbox, I am getting 280+tps (batch and ubatch at 2048) for PP and 18+tps for TG. I really only need 256k context so it all fits.
If I go down to the Q3_XXS quant + dpsark + 'llama-vulkan-radv-performance', I can get about the same PP and 25+tps for TG with draft set to 2 or 3. Fits about the same as above.
Edit: I did notice the 25+tps quickly degrades down to 20+ after the first few hundred tokens.
harness: omp.sh
i used for work where i did less and it quickly reaches thousands if you're not careful. i can already see what some will say: skill issue et cetera - whatever.
> Not everything has to be done by the most expensive one.
Ok then.
$5/days is ~330 Mtok/day, that’s a nontrivial amount of work, and none of the gpts are more efficient than deepseek at $/task if deepseek meets your quality bar.
Not true. Sol on XHigh or Max runs out even on the $200/mo plan. It's not close to effectively unlimited. Maybe at 2x the current allowance it can.
Real work. $200 looks good on the outside until the essence of it, e.g. the models lie. I gave a list of spec to Sol and Sol decided some items didn't need to be done and the reason was "unproven", "not enough evidence", etc.
They all come up with amazing ways to lie (or be lazy). Often times what you get isn't what you asked for (only on the surface). E.g. I ran it to iteratively bench and optimize a better data structure for the project. It spent hours and finally came up with something. When I check it out -- it benchmarked the wrong criteria and was way off. So here we go again. Most AI work looks good on the surface. There are infinite edge cases.
So to do real work and gate it you need to:
1. Plan
2. Get it to do the work
3. Get independent agents to check from different angles
4. Take that feedback and get it to fix those gaps
5. Match against the plan and redo parts if needed
Every task is easily 4-5x the estimated amount of tokens.
p.s. well I did burn some banked resets building a compiler for some language AND it is still NOT done. Every time it says done I say check it says ok we still have bugs...
With 5 active sessions going nonstop? That seems like a pretty important qualifier.
I can pretty easily burn through my weekly quota over several agent coding hours with minimal supervision when tasked with some pretty large but well-planned refactors.
I'm also creating a free platform that replaces extremely out-of-date software, some of it only available with mutli-million dollar contracts, to help medical physics professionals with cutting-edge radiotherapy devices used to treat cancer.
https://brynnbateman.com/ for a list of projects
And out of curiosity, how do you automate testing the porting in the browser that's actually playable etc? And aren't you a bit scared of hosting and serving the "hairy bits" such as full assets? Nice job anyway!
I automate testing playable parts in browser by adding a dev mode that allows text commands for everything instead of having to rely on clicking UI or 3D elements. It can see the full game state in JSON and interact in any way via commands.
And re: IP - I just accept that I might get a C&D any day and have to take it all down. I'm careful to not accept a single penny for any reason and don't even have Patreon. Usually monetizing is what makes IP owners unhappy. And for the Pokemon MMO I just don't advertise it anywhere meaningful since they'll C&D the second they see it. Largely made it for my nephew and we play it together.
Then, once I go over, API pricing racks up FAST!
And to be frank, it is not that much weaker for regular software development work. I use Claude at work and I see no difference in capability. I only notice a dramatic difference in how much more expensive it is.
> I'm running it in Oh My Pi with a second instance running as "advisor" and even with 5-6 active sessions (effectively 12 streams)
OpenCode currently offers 60 USD API credits at 10 USD per month (OpenCode Go) and have even doubled it temporarily as a promotion.
Effectively you can get Deepseek for 1/12th the already ridiculous cheap API price.
Per the open code zen pricing page[1], it appears that the token prices are the same, but their cache is 10x more expensive?
Deepseek 0.14/0.28/0.0028
Pro Opencode 1.74/3.48/0.145
Deepseek 0.44/0.87/0.0036
For flash the input/output is the same, but the cache difference is big, you're paying 10x on >95% of your tokens.
For Pro, it's even worse, input/output is 4x and cache is 40x. The price different is really brutal. Yes you will still come out ahead by spending your first 10$/month on opencode go, but you will be saving a lot less than initially appears from their (60 USD for 10 USD pitch).
[0]: Deepseek: https://api-docs.deepseek.com/quick_start/pricing/ [1]: Opencode: https://opencode.ai/docs/zen/#pricing
The analogy I like is that building software is running a Michelin restaurant. The moment you scale, the chef is just writing cooking books and is absent, and you move into franchising, you will be amazed at the bottom line revenue scaling, while customers will be progressively appalled with the food...
There's no way large companies outside the US will pay the "US AI lab" premium if they can get the same workloads done at a fraction of the cost using open-weight models that they can self-host and optimize/fine-tune on.
Such as Decision Making. /s
You just can't set a high enough threshold of intellectual effort for critical decisions.
Many devs who have never tried from either side build it all up in their head but it's almost always been a matter of thorough tedium, which LLMs are excellent at churning through, especially when there's api docs/headers/code comments.
If you have the space, try mirroring your port at the switch level and capturing every packet then making it go through them all to look for whatever. We have NSA at home lol
For simple queries, we have reached the threshold since the beginning of the year, and models are good enough from every provider to make a meaningful difference between one another. (ChatGPT, Claude, Gemini, Grok, MuseSpark, Kimi, DeepSeek, GLM...)
The real unlock will be, and you can already see it with GPT-5.6 and Fable-5, to delegate complex enough tasks that will take more than 24 hours to get done and they will not lose track. I'm not talking about a loop, but the actual intelligence to recover from these compounding errors that accumulate in dumber models.
We're still a long way from the intelligence needed to let one of these agents go ahead and supervise multiple layers of sub-agents underneath to do complex orchestration. The future looks very promising and exciting. Imagine having the possibility of a Frontier model orchestrating as many sub-agents as needed that are running on cheaper models like DeepSeek.
While SOTAs handle these errors better, they compound in all models and there's a term for that. It starts with cluster and ends with an expletive.
I wish I could, but I don't see the need for human steering going away soon if the task involves anything novel (see Terry Tao's chat).
Yes it's much easier to have a smarter model that goes straight to the correct answer first, but it may not be necessary or economical. There's a minimum bar for the model where it understands problems and knows the right step to correct them, and above that newer models give diminishing returns.
That's basically ASI not AGI, if you agree humans are NGI (natural general intelligence) and make mistakes and wrong decisions in solutions all the time. Right steps with some wrong ones is acceptable though for AGI.
For this genre of task execution can run with limited horizon and is independent but would be too expensive to do with "us frontier tokens", I think for these, there is value in availability of cheaper tokens.
You use Fable 5 right? If that’s good enough for you now, why wouldn’t a Chinese model that’s as good as Fable 5 but at 10% the cost be good enough in 6 months?
I use Claude Code semi-heavily for my small business, and the $100/mo I pay for that is a rounding error compared to the value it provides.
If I can avoid spending an hour or two "massaging" the output from a lower-end model once, or it avoids introducing one load-bearing (sorry, couldn't resist) bug, then that's the entire $100 right there.
Hell, you could argue that the best "coding model" that we have at the moment is the human brain, and people will gladly pay $10,000/mo for one of them.
Arguing over $20 vs $100 for something that actually puts in work just seems insane to me.
Fable 5 is still going to mess things up at any sufficient complexity. The advantage of low cost models with "good enough" intelligence is they can recursively correct. Why? Because it is cheap. Proper requirements and tests and subagents take away increasing amounts of work, at a cost that is not prohibitive.
If you are reviewing code manually you might consider Fable 5 a worse option. As it articulates itself with higher confidence and you already know it is capable, you are may be more likely to miss a mistake. You know to be on guard with a junior engineer. Reviewing a senior who suddenly makes some weird stochastic mistake can be a lot harder. It would be like if the smartest human engineer you knew was capable of some random brainfart in the middle of their massive diff. Imo, much harder to deal with.
Of course, we should keep in mind Fable 5 is only expensive today. It will be cheaper in the future. Autonomous, recursive prompting and improvement is the clear end state. Especially for entities that will always have the budget for that at the SOTA frontier.
Which was an argument for using every less powerful model since the moment they got useful, right?
When was that? Opus 4.5 maybe? Let's say Opus 4.5 for the sake of the argument. So back then we were like "DeepSeek is not good enough, I need Opus 4.5". Now DeepSeek is better than Opus 4.5. So if Opus 4.5 was good enough back then, DeepSeek is better than that now.
Sure, it's always nicer to have a slightly better model. But the price difference starts mattering a lot more when all the models are already sufficiently good.
We've just spun up our first Hermes agent, with direct API access to our main inventory system and that's expected to find another few grand per month in misallocation/inefficiency.
I wouldn't be surprised if we were doing more like $10k/mo higher in 6-9 months' time.
When you're talking about numbers like this, the fact that one AI is $100/mo and another is $10/mo or $40/mo doesn't matter. They could make GLM-5.2, or any other Opus 4.5-class model free and it still wouldn't make sense to deploy in a commercial context.
The other angle I'd approach things from is that Opus 4.5 (and I'd agree with you that that model was the saddle point) was "good enough" for the types of things we were asking it to do back then, but as the models have become more capable the tasks we're asking them to do have also expanded with it.
I know I've personally gone from "hey can fix this race condition with a Redis mutex" 6 months ago to "Independently redesign this full embedded USB stack and QA it end-to-end, working around a specific Kernel bug in macOS Tahoe that requires decompilation to find the source of, while keeping in mind the constraints of our 8-bit AVR chip from 2011" now.
But that said, yes, maybe in 5 years' time we will reach an "intelligence saturation" where the average person won't be able to even conceive of how to use the new SOTA.
Surely they are using read-only access.
This is such a ridiculous objection for how beloved it is. Wide swathes of the public can't cope with any adversity or risk.
With an agent (especially incompetently employed), the danger of unwittingly destroying your company (or at least, the crucial data/reputation) is rather higher. We are notoriously bad at estimating the downside risks in complex systems.
The most obvious case is the downside risks in complex financial constructs... things look great for a while ... until a sudden surprising collapse arrives and totally destroys all the upside you think you have created.
In any case, GP's post was not such a balanced consideration; it was just parroting a beloved risk-aversion meme that can easily be deployed against building anything (what if the building falls on top of someone?) or even leaving home to go to work ("travelling in a hunk of steel at lethal speeds – let me assure you that absolutely nothing can go wrong here, mate.")
What I find tiresome about that meme is the presumption that "something can go wrong" is useful input on its own. It's not. Mistakes are made all the time, the only way to avoid that is to stop breathing. Even in the process of me standing up and going to the loo, something can go wrong.
If the guy wants to make a case that it's too dangerous for the expected benefits, he has to actually make that case. Saying "risk exists" with no elaboration is a waste of HTML. "something can go wrong" every time he swallows food, yet mysteriously he still does it.
(the suicide analogies may seem mean-spirited, but I kind of mean it. If you consider every action primarily from a standpoint of "what harm or irreversible change can result from this", the only permissible path is to do nothing. To be moral is to be as close as possible to a rock or another inanimate object.)
Are you saying we shouldn't care about the future of affordability and access because at this moment we have seemingly endless access?
Sounds extremely short sighted.
If $100 Claud Max subscription works for you, then great.
But you have to remember your pricing is subsidized by enterprises that pay hundreds of thousands of dollars each month, if not more, to Anthropic.
For those companies, a Chinese model that can cut their AI spend from $1M/month to $200k suddenly seems attractive.
And unfortunately for the American tech industry, the valuation is based off those enterprise deals, not your $100/month Claude Max subscription.
This is made brutally obvious by anthropics customer support for people with such accounts.
Right now the US dominates everyone else in actual chips in data centers. So even if deepseek etc tries to undercut, they’re very capacity limited.
It's that good. They are far from capacity limited, and even if they were, you can rent a single MI300X from somewhere like Hot Aisle and get more tk/s than you'll be able to use.
That one is dirt cheap at API pricing, I can't imagine quota is going to be a concern on the $200 subscription, which in my opinion easily supports full time use of 5.6 Sol on xhigh.
A couple of very talented friends were uttering curses upon the entire bloodline of whoever convinced them to try letting sol xhigh do serious work. Deepseek cleaned it up for a fraction of $20.
The cost per task was $0.03 with DeepSeek, $0.05 with Luna. $1.23 for Sol.
Tokens per second 132, 202, 70 respectively.
They also previously said prices will go down significantly once they get a hold of the upcoming Huawei chips (later this year).
Prices are going up just because they can. It can easily come back down. They aren't strained by some IPO / VCs requiring them to 1000x their earnings.
Think about it this way.
Let’s say you could buy an LLM that gets things right 98% of the time. But there’s another LLM that’s 100x the price but gets things right 99.9% of the time. To the lay person this sounds trivial but to a serious business this intelligence gap could represent millions, or billions of dollars.
But that’s simply not the case. It’s very clear that vast majority of the business do not generate additional value from incremental intelligence gain from these models.
There is a reason why Chinese open weight models are now popular even in American enterprises, because CTOs realize that they are indeed good enough.
The Chinese models are not good enough for anything other than pair programming, which is just a very last-gen way of using agents.
And when the big US models get better we will move with them. Until we stop seeing returns there is no "good enough", I don't know why this is so hard for HN to understand.
And even it isn't "enough". I can very clearly see myself using more advanced agents to move up the abstraction ladder.
For businesses that have actual problems to solve, I see them investing in the frontier for a good bit longer, probably until we have AGI that can replace employees, maybe even a bit after.
This is why I find the "good enough" arguments silly. Like, the usefulness of an AI tops out to you when you can pair program with it? Seriously? You cannot envision ways in which more advanced AI enables you to do more, better? That's bizarre to me. I don't ever see myself running out of problems to solve.
Is this what you are looking forward to?
This matches my experience with DeepSeek V4 Pro at Max reasoning, the preview version of the model kept regularly messing things up. About 30-60% of additional time to fix the output was needed.
On similar tasks, GLM 5.2 at Max reasoning screwed up maybe 20-30% of the time, while it still definitely made noticeable mistakes, they were far fewer in total and less egregious.
Kimi K3 at Max reasoning drops that value to below 10%, it's about as good as Opus or approaches Fable in some tasks. At High reasoning it also seems to be pretty close to Opus 4.8, not sure about the latest Opus model yet, but it's up there.
Only problem is that K3 is nowhere near as cheap as DeepSeek models, despite me personally liking the writing tone more (less Anthropic slop) and finding that it doesn't block my cybersecurity prompts, recently reproduced SQLi with a proof of context so I could justify fixing it.
My overall thoughts (released over some time):
https://blog.kronis.dev/blog/ai-slop-is-a-self-inflicted-tra...
https://blog.kronis.dev/blog/kimi-k3-is-out-is-anthropic-don...
https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...
I'd say as Chinese models get better, whatever moat Anthropic and OpenAI have dissipates. Currently the main things keeping me with Anthropic are their performance (tokens/second) and the fact that their visualization abilities within the app are pretty good.
That was ages ago (in LLM release timelines). DeepSeek V4 Flash beats it now and a lot cheaper.
> On similar tasks, GLM 5.2 at Max reasoning screwed up maybe 20-30% of the time,
GLM 5.3 bridges this gap.
> I'd say as Chinese models get better, whatever moat Anthropic and OpenAI have dissipates.
Their moat, especially OpenAI is funding and hardware resources. They gain train models 10x as large and also serve at large scale. That's it.
I’m sure the next models will only get better, when they’re released. Also super curious about what Moonshot will achieve and the full DeepSeek V4 Pro release!
> Their moat, especially OpenAI is funding and hardware resources. They gain train models 10x as large and also serve at large scale. That's it.
I’ve seen how much slower Kimi K3 can be and that part seems correct, their own GPU production still has ways to go and export restrictions definitely limit what they can do.
Not sure about the size part, if Kimi K3 achieves SOTA performance at 2.8T parameters, western models being >2x that size would be insanely bad in regards to efficiency. I bet they’re all within the same order of magnitude and below 10T and won’t really have a reason to go even that high for the foreseeable future.
As investors will start squeezing them for profitability, I suspect focusing more on efficiency will be commonplace.
They're a lot larger e.g. Fable. It is insanely bad. Do you know how much more resources "Western" companies have? Most in China don't have random GPUs to "play with" like every "frontier lab" employee does.
> As investors will start squeezing them for profitability, I suspect focusing more on efficiency will be commonplace.
They're born lucky though. Efficiency is "free". The next generation hardware e.g. Nvidia claims Blackwell -> Rubin is 10x efficiency (verified by Neoclouds apparently).
I do have a way of correcting through redundancy, though. If you are just vibe coding, you need to use the most capable model you can find and even then it might not be good enough.
I mean, this proves my point. Better models enable you to get more done. With Fable, 80% of the time, I no longer have chat with an agent over the details of a PR. I give it an outcome and it gets done. This means I can work on much more with the limited time I have.
And I don't see this ending. When better models come out that take that from 80% to 99.x%, I will have that better model manage teams of other models and move up the abstraction layer.
If models get even better than that, perhaps I stop reviewing PRs entirely. Maybe normies can start using agents to build real things.
Unless your business doesn't have many problems to solve and isn't in a competitive environment, it will benefit from using the best models.
My point is you don’t need the best model if you just put in QA processes that can be done by models also. And if you don’t have that, the model is probably not going to be good enough.
Fable has only been out for a month but somehow everyone is supposed to have moved to a completely different way of working that supposedly only works for Fable and nothing else…
This kind of takes makes me cringe. Why don't you go back to LinkedIn?
I have no idea what you mean by “pair program with an agent”, but Opus have been able of autonomous coding since last November, and with any half-decent harness even local Qwen3.5 was able to do so 6 months ago.
Fable is a stronger model, which means it can solve harder tasks but it's also over-hyped, because only a small fraction of task is hard enough to be Fable-worthy.
Fable is the only one that reliably one shots complex changes and makes the right design choices. Everything else requires handholding.
I can let Fable loose on a 12+ hour (for AI) task and it will have performed it flawlessly when I come back the next day. K3 and Opus are not like this.
And no, our harness is not the limiting factor here.
This is not appropriate for HN. Please review the guidelines: https://news.ycombinator.com/newsguidelines.html
There are businesses other than FAANG. I don't know why this is so hard for FAANG employees to understand.
It's been 2 months since Fable was released to the general, man. Nobody knows what's going on inside of these companies except the people at the coal face.
It's funny to see that Anthopic shills have been saying the exact same thing for the past two years now (and it was OpenAI fans before). It's amazing to see that Claude 3 Sonnet was "great" but now that even Qwen 9B is better than this version of Sonnet DeepSeek V4 is still not good enough despite being stronger than Opus 4.7 was.
> Let’s say you could buy an LLM that gets things right 98% of the time. But there’s another LLM that’s 100x the price but gets things right 99.9% of the time
If you think Fable makes 20 times fewer mistakes than DS4 you're delusional. It doesn't even do 20 fewer mistake than Gemma 4…
Same reason it makes sense to assign a team of humans that cost $100k/mo to a product that brings in $5M/mo, rather than one human with 5 Claude Max subs.
The cost is a rounding error.
This is why I quite like Kimi K3 - close to the same performance (definitely like Opus, approaching Fable), noticeably cheaper, generally good enough for me to daily drive. Only problem is that their official provider (on the Vivace plan) feels kinda slow, I'd say close to 2x slower than Opus on Max reasoning on average (probably more relatable than Fable).
You can say this about literally every product we buy. And yet...
So, death sentence even to frontier models?
Another one I did was a printer data stream translator from an obscure format to PostScript/PDF (or just PNGs), complete with cups support, etc so these old apps can easily be hooked up.
Flash is capable now of running long range defined-goal tasks like this.
At this point, I don't even know if its possible
- You describe the breakdown in terms of time but it's more accurately a function of reasoning complexity.
- You seem to assume that no intermediate evaluation is possible.
- Often it is (e.g. the build breaks or tests start failing), allowing for course correction. There's definitely a cost to that but it can still be cost effective if the accuracy is "good enough" and the price difference significant.
- There are numerous tasks that don't require Fable or GPT5.6 level reasoning to improve efficiency by an order of magnitude.
Spirit was broken by oil prices which everyone pays the same for. (There is no cheaper jet fuel alternative).
Not a good comparison to the point of wrong conclusions.
Spirit was broken by oil prices because they target the low end customer with their ticket prices. Oil prices went up and they had no pricing headroom to charge more on tickets so they simply went kaput. Ryanair also suffered a fair bit. Other airlines did (comparatively) fine because they had the ability to increase prices since their customers are less price sensitive.
It’s a classical business lesson that being a “cost-sensitive” vs a “value-sensitive” business (what this tradeoff is called) is a tradeoff. It’s notably recommended that startups don’t target the lower end in prices since you can’t compete on cost with a business that has more economy of scale than you; you have to compete on features. And being a cost-sensitive business means that you are affected much more than other businesses by changes in material/component prices, because a 13-cent increase in the cost of a component matters more the more product you sell, and if you increase prices too much customers will start to wonder if the “budget” brand is really a good value proposition over the mid-end or high-end ones anymore.
Literally everything you're saying doesn't line up with reality. Apple's most popular product today are two lower end (neo and mini).
You make the mistake of thinking your theory defines reality, when in fact reality says quite the opposite here.
I'm saying this because your logic is the classic logic that everyting thinks makes sense until you do it. Everyone can't target the high end, there are limited customers with many choices, making it far harder actually to win.
You target the demand/pain/need regardless of market.
https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...
https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...
https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...
Seems more like a overall industry problem, not limited to low cost carriers
At least here in Germany Aldi isn't even really limited to the poor, it's famously a place where you can run into anyone. Where I used to live in Berlin close to the government district I literally on occasion ran into the chancellor (and her bodyguards). Aspirational shopping where you buy premium goods to pretend to have higher social status honestly seems a bit on its way out. Even middle class people seem to consciously shop more utilitarian now.
Like Europe?
The ByteDance folks are apparently training a mythos level model 10T params apparently. If they do would it still be subsidized at these cheap rates?
The bet is on using AI to gain competitive advantage. You don't win the stock market or make the deadliest drone by switching to the cheap model
Really? How many times a small team has outperformed a much bigger one just because they were "doing it right"?
I have been in software companies where most software produced was bad. Not just the code, the overall design everywhere. So... bad engineers with the most expensive model, or great engineers with cheaper models?
> You don't win the stock market or make the deadliest drone by switching to the cheap model
The question is not "can you win with the best model?", it is "can you not win without the best model?".
In what kind of sad and failed dystopia is this a "saving grace"? For whom?
For Anthropic and OpenAI, presumably. And the rather large economic distortion field around them, that may or may not go very badly for all our retirement funds if those firms become insolvent...
I've been working with DeepSeek V4 Flash 0731. I'd say that it's maybe not quite as smart as Opus 4.5, but it's willing to think things through carefully and keep going until it gets a good answer. So it's a decent Opus 4.5 replacement. Just let it cook.
It isn't Opus 5 or Fable 5. But it's nearly free on Open Router, and it's self hostable on a Mac Studio with plenty of RAM, or using an RTX Pro 6000 Blackwell or two. Which is chump change for any company that employs programmers.
It would absolutely have been a frontier model last December.
Well... I would think that the whole AI industry in the US are working towards public bailouts... Which I guess they'll get under the current administration... So they'll be fine... Nobody there really seems interested in actually creating a profitable business anyway...
this has already happened with manufacturing, so it isn't surprising that other industries follow.
The US premium in engineering and scientific endeavors have been lacking for the past 30-40 years, and if it werent for tech and silicon valley, the US would have nothing state of the art. Even on that front, the US is falling behind given how much effort in tech has been diverted into privacy invading, and advertising.
The US has been riding momentum, but eventually that momentum will stop. It will take half a century to get back up to speed, and by then, the US will have fallen behind so far that catching back up will seem impossible.
The telling evidence would be if china has the first moonbase before the US does. I think this is highly likely looking at today's US administration.
I joined a company that is an Anthropic shop and I am genuinely shocked.
Sonnet 5 is a little better on long horizon tasks and headless unsupervised agent workflows - but for in-IDE workflows, it's virtually unusable.
I am so used to flipping around my codebase at warp speed with DeepSeek flash. It's so fast and accurate I don't have time for parallel agents. It's a really rewarding workflow.
Moving to Sonnet, you ask is something simple like "split this into a seperate file" "implement this method" "this is my schema, implement a repository for it". It'll spend 30 minutes thinking and charge like $12. And no token caching, what are you even doing Anthropic?
It's unusable.
DeepSeek are in a league of their own
As it is now clear by behaviour of companies and US govt, all these investments will be backstopped by US govt. No US AI company will go hungry, they are national champions.
I dont see how the outlook is any better for the open weight companies. They’re in the exact same situation as the closed weight companies except they have had much less revenue, and built up less of a brand, leading up the the point where they are equal in terms of model quality.
I downgraded my Claude subscription and delegated my Claude Opus access to serve the role of an Architect to brainstorm and plan every step of development.
I leave the development to Deepseek.
Claude gets to review at many layers. It is often just as good as if I let Opus develop it by itself(the architect session will find similar number/level of gaps).
Deepseek flash as an architect and problem solver is not as thorough as Opus5 + high. Codex sol+ high is even better than Opus 5 at this moment for my needs.
First of all, no one knows the "true cost" of any of this, yet, but we know it's expensive. To what extent are the Chinese labs being subsidized? Are they real businesses?
Second, the Chinese labs aren't some "super geniuses", while the American labs are full of clowns. As of today, like the past 3 years, American labs are SOTA. That might change, but let's not act like the American labs don't know what they are doing.
The idea that people are going to use cheaper models for cheaper work isn't some novel revelation, it's completely obvious. People are doing it already, eschewing Fable.
The point is it takes money to keep developing models. Everyone is playing by the same rules. At this point, the US labs are trying to build businesses. I'm not really sure what the Chinese labs goals are. But I do know they aren't doing charity work.
And who gives a flying fuck. I am a "real" business and I count my money. It is not my life goal to prop some fat cats crying crocodile tears. Granted I do not use Chinese models. I use Junie straight from my JetBrain's IDEs that in turn uses Gemini Flash. Very cheap and more than enough for my use.
I have been using it since it got released and its as good for scoped coding tasks, as the other big models I use, but just soo much cheaper.
I’ve found it to be very capable. I’m using it with pi as well and some custom extensions I’ve put together over the past few months and it’s pretty crazy having it do what I need it to a vast majority of the time, do it fast, and see that it’s used like $0.12.
You can probably implement something similar as a plugin for your preferred harness. From a technical perspective I think it just sends the output w/h the thinking and tool trace to another model and asks it to double check everything (exact prompt must be somewhere in the OMP repo).
Would you run a less costly model as the supervisor given it’s consuming a lot of text and may have a simpler task to do like “make sure the implementing model doesn’t start over-engineering things”?
I don't think this is the win you think it is. It's amazing that this is possible, but it introduces so much human overhead that you can drown in reviews and it can effectively slow you down more than a quick check and fix yourself.
The models need to get a lot more consistent in what they can and can't do before you can automate this stuff and only check the things you know the model isn't good at
What's normal usage? I mean, Kimi is already really keen to spin of lots of subagents, and DeepSeep can probably do the same?
> The beauty of intelligence at this cost (even if it's not SOTA) is that it opens a whole bunch of new use cases. Test failure in CI? Have the bot automatically propose a fix, its cheap enough that you can discard it w/h issues.
Yes, though I did that even with Claude (on my employer's token budget). The agents are great at doing the gruntwork of chasing down the reproduction of flaky tests, too. They need some hand holding at first, but the guidelines are usually re-usable per project. (Claude specifically needs to be told to really concentrate on reproduction, and not eagerly start fixing the flake: if you don't have a reliable reproduction, you have no clue whether your fix actually fixes anything.)
Another time, Flash started trying to make tool calls by just calling bash and catting the tool call to stdout. Then it started running echo xx for every two letter UNIX command it could think of: mv, cp, etc and the it dug into uv, ty, and jj
I am not sure why you wouldn't want to use the SOTA models unless speed is a concern. Otherwise you are leaving quality on the table.
The 'floor' has gone up: today's model a bit behind SOTA is like model releases that were blowing people's minds a few months ago. Compared to, say, DS R1, this is far out stuff.
This, Luna, and (if it's good in practice) Laguna S are also fast and light not just cheap. And, as happened before, DeepSeek's first but other open model makers likely follow.
And a small, fast model taking small steps is...fun? More like working with code.
My initial thought was to sign up for ChatGPT, but I had $20 in OpenRouter so I've been trying out DeepSeek V4 Pro with Pi for the last few days and I gotta say, it's good enough for my use case. And even with paying for API usage rather than Claude's subsidised subscription, and with OpenRouter taking their cut, I will probably end up paying significantly less overall. And I really like the flexibility of being able to use whatever minimalist open source harness I want (and being able to switch providers easily, too).
(My demands probably aren't as high as many others' - I mostly use it for help with some hobbyist coding projects, and I tend to ask it questions about how to approach problems rather than just telling it to go off and code stuff for me.)
If you prefer subscriptions, OpenCode Go ($10/mo), Cline Pass ($10/mo), Atlas Code ($20/mo), and CommandCode ($1/mo) serve some of the best open weights with generous limits. OpenCode Go currently offers $120 for $10 on DeepSeek Flash v4 (if you're okay with data retention).
> DeepSeek V4 Flash: ZDR agreement is renewed monthly. The current agreement is valid through August 31, 2026.
Is there other info I should be aware of w.r.t data retention with opencode go? It's hosted in China, so other middlemen may be active (I doubt it, but possible)?
What are you going to do? Take a CCP company in front of a CCP judge?
If they are doing it to you, they are probably doing it to others, which makes an easy class action
(not commenting on whether it was a fair amount - just saying these companies are not immune to lawsuits)
It would be a hassle, but China does have privacy laws. Companies do get sued for violating them.[0]
0. https://www.chinajusticeobserver.com/a/china%E2%80%99s-top-c...
Remember China is still an authoritarian dictatorship. One leader with absolute power for life. Don't let the facade misguide you.
Trying suing when it's in the states interest, like it is to build AI datasets on western data. For reference, no case in China has ever been ruled against the government. They don't have things like judges smacking down executive orders or refunding tariffs.
At DeepSeek's absurdly low rates or market rates?
Careful with using OpenCode's accounting for DeepSeek v4 Pro, though: https://github.com/anomalyco/opencode/issues/39822
I've been running this model locally for a week, and the preview version before that. This updated one feels like a whole tier up. It's very capable for debugging and analyzing documents/data I upload.
The killer feature, IMO, is the speed. On 2x RTX Pro 6000 Blackwell, its ~8k tok/s prefill and ~250 tok/s on a single stream. I saw 1000 tok/s with ~64 concurrent streams on vLLM.
That's fast enough that you can interactively chat with it without switching tabs while you wait, and its a ~300B (13B active, hence the speed) model so the responses are also very good. It's actually more convenient now for me to direct 95%+ of my day to day usage to my local model, and only use Claude Fable for really big coding tasks.
Until this model was released, I was contemplating spending even more money on hardware to run GLM5.2 (~750B) at reasonable speeds, but I no longer feel that need. This is smart enough, and I think it only gets much better for local models from here.
I wouldn't call 80 t/s slow.
I'm guessing tensor parallelism or similar?
Here's a runbook: https://github.com/local-inference-lab/rtx6kpro/blob/master/...
If the newer builds aren't working, you might try running the old v6 build (based on the eldritch-enlightenment image). gilded-gnosis gave me some problems that I haven't bothered to track down, the old builds are still gonna blow away llama-server performance. And that's before you get hooked on vLLM's PagedAttention and can run multiple sequences without a ton of extra overhead.
For reference, on a 1x B300 it's over 400 tok/s decode on a single stream.
It is strong (not Fable strong though) with a much better “persona” than Opus, and very different blindspots. If you flip between Claude and this you will find both catch the mistakes of the other before they get out of control.
On balance I actually prefer DeepSeek for programming now, because of the way it talks.
I suspect that once the hype dies down, or the field gets more competitive we will see the same on 0731.
Spark is actually the interesting one imo. It's significantly better, also significantly faster. If you are ok with letting Meta soak up your data (which DS does too) it's also the same price.
A chinese model being in the same ballpark of capability at half the price sounds believable to me.
DeepSeek just spend almost 2 hours trying to figure out why terrain textures were not working. It tried everything over and over again, it even had reference code for meshes on how to setup the rendering with materials, and it could just not do it.
I finally gave up and gave it to GPT-5.6 Luna instead, and figure out in a single prompt after 20 seconds, that the terrain mesh was being initialized with None in the material slot.
Other tasks it has managed to figure out at least, but it is significantly slower than GPT-5.6 Luna and it requires a lot more iterations.
(Both were set to high reasoning)
Reading the DS reasoning is wild, it's constantly going in circles. The most minor lack of clarity in your prompt and it will spend ages going back and forth on what you meant. It reasons 5x longer than the preview which makes it really slow now as well. We did a lot of work to nudge it to be decisive and improve our evaluation setup to there's more clarity, and it helped but only marginally.
Ours is a full-stack app one shot test so it includes backend, frontend, design, and QA/testing. It's graded by Opus xhigh and Sol xhigh and the grades are averaged.
DS4 preview would finish in 20 minutes flat on high reasoning and grades 6/10. Luna high gets 9/10 in about 30 minutes. DS4-final is crazy - at high thinking it's taking over an hour and getting ~8 but only had one successful run as I got tired of waiting so long after many early abort/retries trying to debug why thinking was so long. The lowest thinking still takes over 45 minutes, and with thinking off it actually is finally closer to preview in time but actually get's a much more varying result anywhere from incomplete to 6 it seems.
Costs per run DS4 is best but not actually by a lot as it's spending 10x the tokens with all the reasoning and mistakes. It's a very brute force model and I really preferred preview in many ways for how predictably fast it was.
Side note, Spark 1.2 is a nice model for this test, best in frontend design and fastest to get results together, though not nearly as efficient as Luna. Grok scores similarly to Spark but at like $50/run vs the contributor Spark costing $1.50.
Edit: was curious to see and seems DeepSWE agrees at least: https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5...
Edit 2: btw it tests a team of agents working together in a special harness and stack. So 20 minutes is for 8 agents essentially. That said everything was built around DS as it was the cheapest to iterate against so even with that advantage the new one struggles.
it's still $3/$15 for all providers on openrouter
because of some Kimi license
Uptime looks crap, though.
So we won't see any price decrease unless Kimi changes the license of K3
Seems like we've reached the event horizon of whether AI advances are worth paying attention to.
Not for me, Fable refuses to debug Linux kernel bugs. Unless you say who you're speaking for, it sounds like you're just shilling for Anthropic.
I use Sol and Grok 4.5 as my inline debuggers/reviewers, and both do well, and are decent at token save. DeepSeek V4 Flash 0731 found some interesting bugs when I tried it a few days ago, and I'm curious to see if that also joins the code-review line up
I recommend opencode or something akin to it to play with models. Any big model updates or hot new ones will naturally run across your desk that way
My read is, OpenAI is neither able to claw b2b money (away from Ant) nor are they able to stave off open weights on the other. In short, they're struggling to hold onto their distant #2 position in the coding market, and these pricing changes reflect a (desperate) change in strategy.
the private endpoint costs 10x (azure).
private endpoints for deepseek (lots of providers) also cost about 10x more.
but 10x more for deepseek is $0.028 cached input, and 10x more for luna is $0.10.
This is on Pi agent, nothing fancy at all about my prompts or use case. Anyone else experiencing this?
I've also had it randomly go from talking about Rust to talking about the electric chair, controversies about D&D rules (both irrelevant and something I've never discussed) and it's completely blind to it in future prompts even when its pointed out and referenced directly
All this said its still worth it but the agentic performance has degraded in my experience at least
When it was first available in opencode, it was kinda slow for me, I guess because everyone wanted to try the new shiny. But now it's back to being screamingly fast and Opus 4.8 level of smart, for penies.
It might be even better in Codex or Oh My Pi according to this bench I saw earlier: https://nitter.net/composio/status/2085330847951970801
Baseten.co's version got into a loop rather rapidly... I've since added loop detection and adjusted some other settings on the pi coding agent and have yet to notice it again. I also switched to DeepInfra ... who serves an fp4 version admittedly, but I've had no issues with it as of yet and it's the top provider on openrouter.ai volume wise.
Which would put them... exactly where everyone else is on this graph.
Edit: I seem to have misunderstood the news. I thought the magical cache read pricing was going away (0.002) and they were going to be on par with everyone else (0.02). But I have no idea.
Edit 2: Apparently, neither do they!
>We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice.
https://openrouter.ai/deepseek/deepseek-v4-flash-0731#provid...
Sort by cache read.
No.
They sent an email to customers saying that they will raise prices "significantly".
How much that will be is speculation.
My guess is that they will just remove the 75% discount they gave when they released V4 preview. It will still be relatively cheap even at 4x the current price.
And how this has been accelerating!!
I felt this very hard when I had to travel in the middle of nowhere in south america, with no network, and wanted to keep an LLM model on my macbook pro with 48GB of RAM. That was back in April 2026, a few months ago.
I downloaded Google Gemma 4 (google/gemma-4-26b-a4b) and - Oh boy - I was amazed by it's capacity!
I was able to use it to code simple things, ask it about nature, learn new stuff while traveling and make stories for the kids.
Was really amazing to observe and experiment this!
Seems to me there will be some good chance to run these great LLM locally on our hardware!
Amazing time to be alive
I Compared Deepseek V4 Flash 0731 (low) to Gemini 3.5 Flash Lite (minimal) and GPT 5.6 Luna (no reasoning) and Deepseek V4 Flash 0731 gets it wrong alot, where as Gemini and 5.6 Luna just gets it done.
Not sure why I was downvoted. But seems the downvoter is quick to downvote anything that doesn’t fit the narrative they’re looking for. I’m just reporting my findings.
But even in this very post, you can see that Max was actually cheaper than High.
If you are using API, you should be comparing based on end-to-end cost or speed or whatever blend of those two matches your cost/time budget.
Someone else here said we could get the model via OpenCode Go for $10/mo and get about $120 worth of credit, so I decided to give it a whirl. My first month is actually $5.
In the 3 hours I've been using it I've burned 3% of my 5-hr, 1% of my weekly and 0% of my monthly.
It's fixed 3 or 4 issues in my C++ game, despite not having visual capabilities to see the screenshots I was trying to give it. One-shotted them too.
Luna struggled with what I thought was an easy task (had to replace a few ASCII chars with the correct unicode char but kept choosing incorrectly).
I'll keep using it.
I have £20/month Gemini and £20 a month claude for a bunch of personal projects.
Yes I have to wait sometimes, it's probably a good thing.
From here on, it's going to become all about harnesses that best situate and organize swarm intelligence at scale.
Furthermore, in my company we are using MCPs for Google Ads (it manages our ads), Analytics, Search Console, Zoho CRM, Microsoft Clarity... We use it to crawl specific websites and send daily summaries to our sales team in MS Teams channel. We use it to send daily summaries on marketing statistics and analytics... All with a FREE model. We are rarely hitting any limits so far and in case we need more tokens - we use NOUS or openrouter to pick between Flash or Pro for specific tasks that require more churning.
AMA.
[0] https://taylor.town/silver-landmines
When I see dramatic leaps like this, it tells me that the important hacks haven't yet been discovered.
But note that you have to use Cline (or other harness) if using vscode. I was shocked at how poor the recent versions of GitHub Copilot are at using the cache (with Fireworks AI, but I believe it's a more generic problem).
And a meaningful chunk of the comments are saying "this piece of garbage isn’t even at the level of gpt-oss 20B".
For anything even moderately complex.. like, even low end of complexity, this model behaves maximum like gpt-5.6-luna-high .. nothing more.
Yesterday itself I gave it a coding task in some existing moderately complex small project, and i was using xhigh thinking effort, it was unable to cover all edge cases... and i had already got it to review, and then fix, 3 more times, after the first initial one.
Still it left 2 edge cases.
Then, reverted full code, gave sol-high the same task, it took well over 20 minutes, and completed it in one go with zero edge cases remaining.
I am not using it for anything serious anymore.
I am among those with real life experience with the model that used the previous as well and will attest that the new model is a big improvement
Tell your PjM who should tell your PgM who should tell your PdM, all the PMs...
Maybe if "the business" sees it is true of LLMs, they might believe it's true of giving better context to engineers up front then giving them time to think and prototype (thinking tokens are an answer prototype).
If I'm reading the chart correctly, a couple observations:
* deepseek-v4-flash-0731 max is better than kimi-k3 max
* glm-5.2 is dumber than a box of rocks (this must be on low reasoning or something, right?)
This is way more extreme than other results I'm seeing, like those from Artificial Analysis.
Does no thinking emissions for context saving.
Btw if you need an app I may deliver it to you in ten minutes for just five cents if I'm in the mood. Just let me know.
ARC-AGI II:
- GPT-5.2 (medium) %26.7 ($0.759)
- DSV4-Flash (max) %61.4 ($0.04)
Not a huge deal since it's still cents per session, but my bigger issue was the weird change in tone. It became a lot more pretentious and over-explanatory.
Heavy prompt reworking helped but maybe that's just the cost of being better at coding and ARC-AGI?
When I need vision capabilities I use GPT 5.3 codex and if deepseek can’t figure something out after a few goes I switch to GTP 5.5 or 5.6 (I’ve been giving Terra first bite recently and it does pretty well, and have used Sol a couple of times).
Using this regimen means I spend under $100 per month on inference and I work all day everyday with multiple agents running simultaneously all on API token spend not subscriptions.
What secret sauce do they have?
pair it with codewhale, 50 agents, 200 MB of ram.
Quant company usually squeezing every penny.
It’s always a bit tricky picking the right harness (when you have options). Sometimes the differences are subtle but meaningful. But who has the time to run everything twice and compare all the time!
Codex is really good in my experience, especially due to its native sandboxing. Deepseek seems really well versed in its tools, including update_plan and knowing when to request sandbox escalation.
https://reddit.com/r/DeepSeek is where the fellow F5ers are at.
But it makes me quite curious, how a text-only model can do so well on ARC-AGI-2 being a set of visual puzzles? It would have to solve it entirely using text-only spatial reasoning about the grid (or maybe writing code?). I am curious if this is normal or do other models use their vision capabilities to solve the puzzles?
China has zero energy concerns in terms of energy production - not literally zero, but they’d be able to prioritize other dimensions and not necessarily worry about efficiency
Here they are though releasing models that sip resources
promising!
https://americanliterature.com/author/em-forster/novella/the...
That's irrelevant when you use $/task as the metric, which the OP does use.
That's how I handle the Qwen27B and 35B
What do you mean by "redirect it to useful output"? Could you give an example? This sounds interesting.
But I find it having a pretty significant problem with tool calling - no idea why, but tool calling with it is SLOW. As long as the model is reasoning, all good. But give it a bunch of tools and it becomes extremely slow.
Am I the only one experiencing this?
Imagine if they had GPU resources of western labs.
SV companies get way too comfortable when they have enough in the bank to stay running more than three months.
Energy and intelligence are good too, sure.
As an end consumer, I don't care about the number of active parameters. I really do care only about the tracked metric (how well does it do the job, and how much does it cost... ideally also with time included, but that wouldn't fit on a 2D chart)
With the exception of cache costs, all providers have similar input/output costs.
They don't strictly need any kind of subsidies.
FWIW they have a funding round planned (kerfuffle about leaks from CEO presentation few weeks back) -- presumably because infrastructure needs have ballooned.
Naturally there will be some PRC government interest in one of their flagship AI companies. From what is visible seems to be more along the lines of ensuring that DS gets its fair share of resources -- e.g. Xi Jinping meeting founder and positive comments about success of DS means that (hypothetically) Alibaba can't screw DS too much on infra charges to kill off a 'competitor'. Also would imagine that DS's top guys have been clearly identified and will have been 'discouraged' from going to work for one of the SV polycules. But even here as much carrot as stick -- none of the DS top guys will ever need to work again except for love of the job.
> They don't strictly need any kind of subsidies.
You understand how these two sentences directly contradict each-other, yeah? The money-losing operating of training a model is paid for by momey earned from prior investments. So… the work is “subsidized” by its parent company’s investments in it.