Technically DS is an independent sibling of Citroën within Stellantis, a sprawling car conglomerate that owns a dog’s dinner of car brands in Europe and USA.
If we want to get really technical, “DS4” is a model from Citroën and they later spun out the DS lineup into its separate brand, with the “Citroën DS4” becoming “DS 4”, “DS” being the make and “4” being the model.
Not really. I'm building now another fast C compiler with DeepSeek 4 Flash, and rarely have to step outside to use Pro or Sonnet, gpt or kimi-2.6. Flash is very capable of almost everything.
> The blog post implies that it currently requires 96GB of VRAM.
Has anyone tested what happens if you try and run this on lower-RAM Macs? It might work and just be a bit slower as it falls back on fetching model layers from storage.
True (with 64GB RAM it'd have to fetch 20% of its active experts from disk already, about 650MB/tok at 2-bit quant - and that percentage rises quickly as you lower RAM further); my question is just a more practical one about whether it runs at all, how bad the slowdown is, and to what extent you might be able to get some of that decode throughput back by running multiple (slower) agent sessions in parallel under a single Dwarf Star 4 server.
FYI, llama.cpp (which antirez/ds4 is inspired by) supports system ram. E.g. [1] is a good guide for running a similar-sized model with 128gb ram and a 3090-sized GPU.
Have you had it getting stuck in endless loops maybe ~10-20% of the invocations? Seems it happens for both the responses and chatcompletion APIs, and no matter what inference parameters I try it happens at least for 1/10 of the requests, I've tried every compatible vLLM version + currently using it from git (#main) yet the issue persists.
Seems to happen with various quantizations too, even the NVFP4 versions and any others, so seems like a deeper issue to me, or hardware incompatible perhaps.
DeepSeekV4 Pro is really really competent model and what makes it extremely good is the price point it is offered at.
I have been toying with a 2.5D engine in C on on top of raylib and using DeepSeek as companion in between.
It's thinking transcripts in OpenaCode are transparent and mind boggling to look at things it would consider in its thought process. Very long to read but none of them useless or meaningless.
Always happened that I discovered an assumption that I didn't think about or was wrong but DeepSeek flags it in its thought process and then in final output it would "align" to my flawed request and I'll tell it wait, I saw you thought so and so too and that's correct I made a mistake let's consider that aspect too.
I've been using the Q4 version on my Mac Studio over my local network and it's been good. Indeed, I had the first ever experience where I was playing with it alongside my various other agents and forgot it was a local model as it was doing such a good job.
I do wonder, though, if another agent is really needed. I've been driving it with Pi (Claude Code's system prompt is far too heavy given the prefill speeds) and it's been great. OpenCode is another good option. Is there anything else to gain from another similar tool specific to Deepseek 4?
There is no need for another agent, functionally. But if you follow the idea of DS4 itself: the API agents use forces to do odd things, like translating the DSML stanzas to JSON, with all the canonicalization / KV cache checkpointing problems resulting from that. Is it really the case? What about also providing a sane alternative? Also I'm not sure why people don't try to write more stuff in that area in C/Go/Rust to have more control / speed / less dependences.
Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes: https://x.com/antirez/status/2055190821373116619 Now that code is cheap, ideas have more value. Are we sure that today it is still the case to think in terms: "Is another XYZ needed"? It could be the case that only just to explore new ideas, it is worth it. I I don't like the Javascript / Node ecosystem for my code, so if I have to explore a new TUI or agent workflow, if I do it with the tools I'm more happy to use, the result, the iterations, are different.
> ...I'm not sure why people don't try to write more stuff in C to have more control / speed / less dependences.
Codex CLI is written in Rust, which should give comparable raw performance to C/C++. Of course you can care about the "less dependencies" point but this is somewhat less of a concern on a properly maintained project like Codex. That's not so much "wild, out of control" third-party dependencies and closer to the old ideal of proper software componentry.
> Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes.
This mockup is really nice and the sidebar display gives you a natural way to expose running multiple thinking flows in parallel, at least if you keep them from stepping on each other's toes with code edits (keep them all in read-only "plan" mode or working on completely separate directories/files). That's not so helpful on a 128GB MacBook where a single agentic flow brings you to thermal/power limits already, but it suddenly becomes useful on other hardware (DGX Spark, Strix Halo, lower-RAM machines with SSD offload, multiple nodes with pipeline parallelism) where you have more compute than you could use for single-stream decode.
Love this, even if can't use it atm (not got the h/w - only 96gb on M2 Max). I get it the general comp/public will find it unusable or worse. Reminds me of how home computers were - mere toys - before they became personal computers (PC). On my h/w the only passable combo for me atm is pi agent + llama.cpp + nemotron cascade-2 model: to 1M context, hybrid arch doesn't crash & burn 1/N^2 with context depths of 10K-50K-100K used by code agents. Was on a plane without Internet the other day. Brought a smile to my face that I could run pi agent (with llama.cpp serving), and it was just about usable at 40-30 tok/s. Afaik the usual API speeds are double that, 60-80 tok/s. Sensors showing using 60W when running inference. So battery probably would not last more than >3h. Model only 30B in size leaves plenty of space for KV-caches, and other programs - even at generous 8-bit quant. Only 3B active params at one time (with MoE A3B) is about the most that ageing M2 Max can carry it seems.
> even if can't use it atm (not got the h/w - only 96gb on M2 Max).
Not sure if it works different on macOS, but with CUDA + DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf I can fit it within 96GB of VRAM, together with context, so theoretically I feel like you should too, unless macOS uses GB of RAM/VRAM for the OS/display by default.
I'm very curious where we will saturate the curve on "enough" intelligence for coding. At some point, you can let a less smart model hammer at a problem for longer and get to the same result, and as long as you are not involved it comes to the same thing. I feel like DeepSeek V4 Pro is nearly there. Maybe Flash is too.
Once we hit that point, I am curious how much of Anthropic's current business model falls apart? So far it's always been clear that you just pay for the most intelligent model you can get because it is worth it. It now seems clear to me that there is limited runway on that concept. It is just a question of how long that runway is. I honestly wonder how much of their frantic push to broaden out into enterprise / productivity is because they see this writing on the wall already.
> At some point, you can let a less smart model hammer at a problem for longer and get to the same result, and as long as you are not involved it comes to the same thing.
Is that true? I find the smarter models can just be effective when smaller models can't. It isn't a matter of just waiting longer.
it's almost certainly not true yet but at some point there might be an equilibrium reached of speed Vs quality (and let's not forget, cost) where it's true for most of what you do.
Perhaps you'd still turn to hosted models for the hardest tasks, but most tasks go local. It does seem like that would make demand go down significantly.
Of course that's all predicated on model advances plateauing, or at least getting increasingly more expensive for incremental improvements, such that local open source models can catch up on that speed/quality/cost curve. But there is a fair amount of evidence that's happening. The models are still getting noticably better, but relative improvement does seem to be slowing, and cost is seemingly only going up.
* local compute isn’t scaling as before, so algorithmic improvements are the only ways models get meaningfully faster and smarter
* all those same algorithmic improvements would also be true for larger models
* hardware manufacturers have an incentive against local LLMs because cloud LLMs are so much more lucrative (+ corps would by desktop variants if they were good enough)
So no it’s not clear quality will ever be comparable. It may be good enough for what you want but there will always be a harder problem that you need to throw more compute and more memory at.
> It may be good enough for what you want but there will always be a harder problem that you need to throw more compute and more memory at.
Sure, but if the “good enough for what you want” consumes the vast majority of cases - data-center ai becomes just for the very extreme edge cases. Like how I can render a 4k rez video game at 60fps on my home pc, but if pixar wants to render their next movie they use data-center compute.
> all those same algorithmic improvements would also be true for larger models
Smaller models run faster. If ten runs of a small model gets me the same quality result as one run of the big model, and the small model runs 10x faster, then they are functionally the same.
> I'm very curious where we will saturate the curve on "enough" intelligence for coding. At some point, you can let a less smart model hammer at a problem for longer and get to the same result, and as long as you are not involved it comes to the same thing. I feel like DeepSeek V4 Pro is nearly there. Maybe Flash is too.
It's always going to be cost;
developer time vs developer cost vs AI cost vs developer productivity.
With 4.6 it's looking like we are at the upper limit of appetite for cost (for "regular" Business) so the other levers will probably need to change.
Kilo (the open source coding agent) tested Deepseek v4 Pro and Flash vs Opus 4.7 and Kimi K2[1].
It did ok, but scored substantially less than Opus. It also cost nearly as much, even with the current launch promo pricing for Deepseek.
That cost is interesting - I've seen similar things with Sonnet vs Opus, and in my own benchmarking there are some models that benchmark well, seem to have a good price but use so many tokens they cost just as much as "more expensive" models.
I imagine we'll get to "good enough" for hobbyist programmers fairly quickly, but businesses will still be willing to pay more for faster and smarter. Why make your programmers wait?
That depends on where the methodology goes. But more and more it's hands off. If the trajectory continues it won't matter because nobody is sitting their waiting / watching the LLM code anyway. It is all happening in the background. We might see hybrid approaches where the weaker / cheaper agent tries to solve it and just "asks for help" from the more expensive agent when it needs it etc.
> nobody is sitting their waiting / watching the LLM code anyway
My personal experience is that for production-grade code you need to steer the agent more often than not... so yes, at least some of us are watching the LLM code.
Metal is our primary target. Starting from MacBooks with 96GB of RAM.
NVIDIA CUDA with special care for the DGX Spark.
AMD ROCm is only supported in the rocm branch. It is kept separate from main
since I (antirez) don't have direct hardware access, so the community rebases
the branch as needed.
> This project would not exist without llama.cpp and GGML, make sure to read the acknowledgements section, a big thank you to Georgi Gerganov and all the other contributors.
Edit: aww, doesn't seem to support offloading to system RAM[0] (yet)
Has anybody tried it?
There is a lot of emphasis on MacBook Pro in this thread, but I would like to use it with an AMD Halo Strix with 128GB of unified RAM.
They removed the baseline 8GB RAM/256GBstorage model. My bet is with increased RAM prices the markup on the lower end is not enough to still make a profit
When I ran DS4 Q2 the other day (without the new update Q2 imatrix) it was behaving quite poorly after a few agentic turns with opencode, it couldn't modify the files it was telling me the work was ready and didn't use any tool to update files
It’s shocking how close this feels to claude, obviously it's much slower, but I don’t know that it’s significantly dumber. Interestingly the imatrix quantization seems to be better than whatever quant the zdr inference backends on open router are using. It was self aware enough yesterday to realize that it’s own server process was itself without me telling it, which is not something I’ve ever observed a local model doing before.
In my (obviously anecdotal) testing, DeepseekV4 Pro was better than Sonnet at coding. However, it is much slower, but also many times cheaper, especially with the promotion right now.
You pay per api call but you will be challenged to burn trough 20$ per month. 24/7 usage for single agent will probably cost you around 100$ per month. It is very efficient especially with modern harnesses.
I don't fully understand why it is a standalone project. The readme discusses this: DwarfStar 4 is a small native inference engine specific for DeepSeek V4 Flash. It is intentionally narrow: ...
I think the only bigger difference in DeepSeek V4 vs other models is maybe the type of self-attention. And that leads to: KV cache is actually a first-class disk citizen.
But I still feel like those changes could have been implemented as part of some of the other local engines.
I also assume more models will come out, not just from DeepSeek but also from others, and they might share similar self-attention approaches, that would benefit from a similar KV cache implementation.
Which makes sense, the amount of PRs llama.cpp receives from authors who have no clue what they're doing and can't even answer simple questions about "what they did" is staggering, must be very exhausting to have to figure out "is it worth replying to this author?" for every single PR.
I don't see an explanation of why they would make a model-specific inference engine vs just using llamacpp. There are already lots of people working on the llamacpp integration. This is a lot of effort spent on a single model which is likely to become obsolete when a different model comes out that does better. In some discussions, people are now making PRs against both the llamacpp branches and ds4... so it's taking a rare commodity (people investing development time in this model) and fragmenting it
way easier to work on a focussed c codebase you own than a mature unwieldy c++ codebase you don't. but it's fine, people will take that work and port to llamacpp and everyone wins.
(the ux of ds4 is fantastic too -- it's dead-easy to get a known-good model, great quant. llamacpp you're much more hacking in the wilderness, w/ many many knobs.)
Author has mentioned many times that the llama.cpp maintainers don't want code that's prevalently written by AI with no human revision. If anyone wants to try and get the support upstreamed into that project, they're quite free to do that: the code is MIT licensed.
Also Antirez has been able to use GPT to iterate on the code and performance. He/they (others contributed to DS4) has a set of result files to ensure that correctness is maintained, and benchmarks to verify performance, and the LLM is able to iterate within that framework. Having a small, focussed codebase helps here.
At a certain point the level of abstraction / genericization necessary for a big flexible project (like llama.cpp or Linux) blows things up into a huge number of files. Something newer and smaller can move faster.
I want something like this but not only for my own computer but also for client projects or stuff I might run in cloud GPUs. Because the core idea of having a strong model that is efficient and doesn't require a cluster still applies to a lot of business cases. I am hoping something like this can work in batch mode.
Right now I feel like a 4bit Qwen 3.6 27B with MTP is one of the best for agentic tool calling for some smart voice agents in an H200. I wonder if DS4 Flash being using 80b at 2 bit with 13b active and MTP could be even faster and smarter and allow more concurrent sequences?
This special 2bit quantization seems like a big deal.
With "intelligence" (or whatever you want to call it) and speed both seeming to ramp up quickly with local models I wonder what the growth rate and ceiling(?) might be in this space. Will this kind of iq and performance work with just e.g: 16GB RAM in a couple years? Is there a new kind of Moore's law to be defined here?
Squeezing a model like this complete with 'big model smell' into 16GB...Honestly it's not even possible or feasibly possible today.
It'll require some kind of:
- breakthrough in architecture or
- breakthrough in hardware or
- some breakthrough quantisization technique
The problem is that all the parameters need to be in memory, even the ones that aren't active (say for Mixture Of Expert Models) because switching parametrs in and out of ram is far too slow.
We show that EMO – a 1B-active, 14B-total-parameter (8-expert active, 128-expert total) MoE trained on 1 trillion tokens – supports selective expert use: for a given task or domain, we can use only a small subset of experts (just 12.5% of total experts) while retaining near full-model performance."
The people working at the leading edge of this stuff seem to believe that there is a need for parallel models that solve different problems.
A crow exhibits some degree of intelligence in what is a very small brain compared to humans. There is overlap in the problem solving skills of the dumbest humans and the smartest crows.
So the question is: what is that? Yann LeCun seems to think it’s what we now call world models. World models predict behaviour as opposed to predicting structured data (like language.)
If your model can predict how some world works (how you define world largely depends on the size of your training data), then in theory it is able to reason about cause and effect.
If you can combine cause and effect reasoning with language, you might get something truly intelligent.
That’s where things seem to be going. Once we have a prototype of that system, there will be many questions about how much data you really need. We’ve seen how even shrinking LLMs with 1-bit quantization can lead to models that exhibit a fairly strong understanding of language.
I don’t think it’s unreasonable to expect to see some very intelligent low (relatively) memory AI systems in the next couple years.
I got this running on a 128GB M5 the other day - pretty painless, model runs in about 80GB of RAM and it seemed to be very capable at writing code and tool execution.
Comparison with a RTX Pro 6000, with DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf:
prefill: 121.76 t/s, generation: 47.85 t/s
Main target seems to be Apple's Metal, so makes sense. Might be fun to see how fast one could make it go though :) The model seems really good too, even though it's in IQ2.
I don't want to be a jerk but 31t/s prefill is basically unusable in an agentic situation. A mere 10k in context and you're sitting there for 5+ minutes before the first token is generated.
So you’re saying I should buy the M5? :) I’ve been resisting, thinking I’ll never use it… it’ll be better in a year… I’ll wait for the Studio (do we still think that’s coming in June?)… etc.
I expect this to be my main machine for the next 3-4 years (which is how I justified the 128GB one). It's a beast of a machine - I love that I can run an 80GB model and still have 48GB left for everything else.
Can't say that it wouldn't be a better idea to spend that cash on tokens from the frontier hosted models though.
I'm an LLM nerd so running local models is worth it from a research perspective.
An M5 Max MBP with 128G of RAM costs ~$5k. An Nvidia RTX 5090 with 32G RAM is $4-5k, and RTX PRO 6000 with 96GB RAM $10k. Do you have any data on which is the best price/performance for local inference? Do you know what the big OpenAI/Anthropic/Google datacenters are running?
As always: it depends on your needs. Here's a very basic heuristics rundown:
- More RAM: bigger models, more intelligence.
- More FLOPs: higher pre-fill (reading large files and long prompts before answering, the so-called "time to first token").
- More RAM bandwidth: higher token generation (speed of output).
So basically Macs (high RAM, okay bandwidth, lowish FLOPs) can run pretty intelligent models at an okay output speed but will take a long time to reply if you give them a lot of context (like code bases). Consumer GPUs have great speed and pre-fill time, but low RAM, so you need multiple if you want to run large intelligent models. Big boy GPUs like the RTX 6000 have everything (which is why they are so expensive).
There are some more nuances like the difference of Metal vs. CUDA, caching, parallelization etc., but the things above should hold true generally.
Some of my colleagues believe that current frontier AIs are too heavily subsidized and it will come to an end. They think frontier coding AI's might get unavailable for one reason or another. But these kind of projects show that with 6000$ Macbook we are getting closer to a local frontier model. More importantly, it shows the genie will not go back into the bottle.
> This project supports steering with single-vector activation directions; [...] This is also useful for cybersecurity researchers who want to reduce a model's willingness to provide dual-use or offensive security guidance.
Wink wink, nudge nudge.
I have a feeling most cybersec researchers would only be interested in negative values of "reduce" :D
> Gentle reminder on how, in the recent DS4 fiesta, not just me but every other contributor found GPT 5.5 able to help immensely and Opus completely useless.
I've noticed the same for lower level squeezing-as-much-performance-as-possible code work.
Assuming we are talking about Code/Codex are you on API billing or subscription? I have essentially unlimited API billing at my disposal and I haven't noticed any degradation of quality across Opus versions.
Same here, the enterprise version of Claude has been great. Luckily I’m not the one paying for it. We also have CoPilot and when GPT-5.4 came out, and was 1x request cost, I was very impressed but haven’t had much time to compare the two.
I also don’t have time to do much personal coding outside of work, so I haven’t subscribed to a personal one yet. But I intend to go for Codex just to balance the Claude at work and also because of the hostile moves from Anthropic toward their consumer business.
There's so much subjectivity with models. As soon as a new model comes out people act like the last model they used for 6 months was completely useless.
There is a benchmark for performance work, and I think it is not being optimized by model vendors. The latest result from GSO is that both Opus 4.6 and 4.7 slightly outperforms GPT 5.5. This also matches my experience.
Just want to mention that I've been pulling down and using DwarfStar locally and it's incredible. I actually have it running on my personal macbook m4 max with 128gb of ram and I am running the server to share it through tailscale with my work laptop and just have pi running there.
The long context reasoning is something I haven't even seen in frontier models - I was running at 124k tokens earlier and it was still just buzzing along with no issues or fatigue.
I am amazed at how well it works, I'm using it right now for some pretty complex frontend work, and it is much much faster than, for example running a dense 27b or 31b model (like qwen or gemma) for me (The benefits of MoE) - but the long context capabilities have been what have been absolutely flooring me.
Super excited about this project and hope Antirez can keep himself from burning out - i've been following the repo pretty closely and there are a ton of PR's flooding in and it seems like he's had to do a lot of filtering out of slop code.
Did someone compare DeepSeek 4 Flash to Qwen3.6-27B on real tasks (quality + speed)? According to the benchmarks at artificialanalysis.ai, Qwen3.6-27B is better at agentic tasks, and DS4 is only 2 points better at coding (both with max reasoning effort, full weights). At the same time, DS4 requires 5 times more VRAM even at 2 bits. Last time I explored this topic, large MoE models at 2-3 bits usually performed worse (quality-wise) than dense ~30B models at 4-8 bits, despite being much heavier to run.
Sure, MoE models have more knowledge, but extreme quantization may negate the benefits. And generally for coding tasks, you don't need a model that has memorized all the irrelevant trivia like, I don't know, the list of all villages in country X. DS4 also seems to run much slower on Mac Studio Ultra, which appears to be more or less in the same price range as RTX 5090. RTX 5090 gives me 50-60 tok/sec and 260k context with Unsloth's 5-bit quantization (only some layers are 5-bit too) and an 8-bit KV cache; prefill is instant too. It works flawlessly in OpenCode.
If you already have a spare high-end Mac, I can see the benefit, but I'm not sure it's a good configuration overall. Unless Qwen3.6 is more benchmaxxed than DS4 :)
It is a big thing for sure to have a competitive local agentic model. I've replaced gemini 3 flash preview with DeepSeek v4 flash for all of my personal use cases. Starting from chat app, language learning, and even hobby coding. For coding, I couldn't get decent results no matter which sota latest models I used before. It's not close to Opus or Codex models. It's a flash model and makes mistakes here and there (I just saw `from opentele while import trace`, new Python syntax!)
But I found its tool calling is reliable than other oss models I tried. I assume that it attributes to interleaved thinking. Its reasoning effort is adjusted automatically by queries. I enjoy reading these reasoning traces from open models because you can't see them from proprietary models.
I would love to try DS4 so bad. Well, I don't have a machine for it. I will just stick to openrouter. I wish I can run a competitive oss model on 32GB machine in 3 years.
> I wish I can run a competitive oss model on 32GB machine in 3 years.
You could try DS4 on that machine anyway and see how gracefully it degrades (assuming that it runs and doesn't just OOM immediately). Experimenting with 36GB/48GB/64GB would also be nice; they might be able to gain some compute throughput back by batching multiple sessions together (though obviously at the expense of speed for any single session).
> I wish I can run a competitive oss model on 32GB machine in 3 years.
It's so hard to predict what size the open-weight models will be, even in 6 months time. Will a 96GB machine turn out to be a complete waste of money? Who knows.
FYI, this to me points to an inference bug, bad sampling, or a non-native quant. OpenRouter is known to route requests to absolutely terrible, borked implementations. A model like DeepSeek V4 Flash shouldn't be making syntax errors like this.
llama.cpp (and consequently LM Studio) don't support DeepSeek V4. If you want to run V4, this is your only option right now unless you have hardware that can run vLLM.
96GB is what the author claims will work in a foolproof way for easy production use. But nothing stops you from trying to run it on 48GB, it ought to gracefully fall back on accessing model layers from the disk.
I'm having trouble finding dictionaries or other references that add the qualifier that it needs to be self-tested and not relying on the research of others. Can you point me to one?
Are you comparing an LLM running on a laptop to a Porsche?
I just find it really funny people are willing to write things like "empirically speaking, X is obvious" without actually testing it themselves.
I've seen mixed reviews, and the most honest sounding ones have said it has latency issues.
I don't really care that much what the average LLM power user says at this point, they're impressed by anything an LLM does. They're like toddler's entertained by the sound their Velcro shoes make.
You LLM people are going to be like my mom, once she got an Maps app she completely gave up on navigating anywhere with her own brain, and is lost without a phone.
Except for you LLM people, its going to be reading, writing, problem solving and thinking in general. You'll be completely reliant on an llm to get anything done, have fun with that. You're cooked bro.
Not even close. "I made this DSP task faster by focusing on exactly one compute graph on one machine instead of a compute graph compiler that runs on every possible machine" is a real engineering approach, and the AI usage is incidental. Things like Gas Town are self-serving turboslop whose only purpose is to generate more slop.
I thought DeepSeek was closed-weights and proprietary? I wonder how it compares against Western open-weight models. The hugging face page contains the comparison only with proprietary models for some reason.
For others who are lacking context :-)
Their stated inspiration for this SEO bomb is Chanel perfumes.
Has anyone tested what happens if you try and run this on lower-RAM Macs? It might work and just be a bit slower as it falls back on fetching model layers from storage.
From the Github page it seems it only supports Apple and DGX Spark. I have 128 GB of RAM and a 3090 but it probably won't work.
[1] https://unsloth.ai/docs/models/tutorials/minimax-m27
(Unsloth's deepseek-v4 support is still WIP)
Seems to happen with various quantizations too, even the NVFP4 versions and any others, so seems like a deeper issue to me, or hardware incompatible perhaps.
I have been toying with a 2.5D engine in C on on top of raylib and using DeepSeek as companion in between.
It's thinking transcripts in OpenaCode are transparent and mind boggling to look at things it would consider in its thought process. Very long to read but none of them useless or meaningless.
Always happened that I discovered an assumption that I didn't think about or was wrong but DeepSeek flags it in its thought process and then in final output it would "align" to my flawed request and I'll tell it wait, I saw you thought so and so too and that's correct I made a mistake let's consider that aspect too.
I do wonder, though, if another agent is really needed. I've been driving it with Pi (Claude Code's system prompt is far too heavy given the prefill speeds) and it's been great. OpenCode is another good option. Is there anything else to gain from another similar tool specific to Deepseek 4?
Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes: https://x.com/antirez/status/2055190821373116619 Now that code is cheap, ideas have more value. Are we sure that today it is still the case to think in terms: "Is another XYZ needed"? It could be the case that only just to explore new ideas, it is worth it. I I don't like the Javascript / Node ecosystem for my code, so if I have to explore a new TUI or agent workflow, if I do it with the tools I'm more happy to use, the result, the iterations, are different.
Codex CLI is written in Rust, which should give comparable raw performance to C/C++. Of course you can care about the "less dependencies" point but this is somewhat less of a concern on a properly maintained project like Codex. That's not so much "wild, out of control" third-party dependencies and closer to the old ideal of proper software componentry.
> Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes.
This mockup is really nice and the sidebar display gives you a natural way to expose running multiple thinking flows in parallel, at least if you keep them from stepping on each other's toes with code edits (keep them all in read-only "plan" mode or working on completely separate directories/files). That's not so helpful on a 128GB MacBook where a single agentic flow brings you to thermal/power limits already, but it suddenly becomes useful on other hardware (DGX Spark, Strix Halo, lower-RAM machines with SSD offload, multiple nodes with pipeline parallelism) where you have more compute than you could use for single-stream decode.
Not sure if it works different on macOS, but with CUDA + DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf I can fit it within 96GB of VRAM, together with context, so theoretically I feel like you should too, unless macOS uses GB of RAM/VRAM for the OS/display by default.
Once we hit that point, I am curious how much of Anthropic's current business model falls apart? So far it's always been clear that you just pay for the most intelligent model you can get because it is worth it. It now seems clear to me that there is limited runway on that concept. It is just a question of how long that runway is. I honestly wonder how much of their frantic push to broaden out into enterprise / productivity is because they see this writing on the wall already.
Is that true? I find the smarter models can just be effective when smaller models can't. It isn't a matter of just waiting longer.
Perhaps you'd still turn to hosted models for the hardest tasks, but most tasks go local. It does seem like that would make demand go down significantly.
Of course that's all predicated on model advances plateauing, or at least getting increasingly more expensive for incremental improvements, such that local open source models can catch up on that speed/quality/cost curve. But there is a fair amount of evidence that's happening. The models are still getting noticably better, but relative improvement does seem to be slowing, and cost is seemingly only going up.
* local compute isn’t scaling as before, so algorithmic improvements are the only ways models get meaningfully faster and smarter
* all those same algorithmic improvements would also be true for larger models
* hardware manufacturers have an incentive against local LLMs because cloud LLMs are so much more lucrative (+ corps would by desktop variants if they were good enough)
So no it’s not clear quality will ever be comparable. It may be good enough for what you want but there will always be a harder problem that you need to throw more compute and more memory at.
Sure, but if the “good enough for what you want” consumes the vast majority of cases - data-center ai becomes just for the very extreme edge cases. Like how I can render a 4k rez video game at 60fps on my home pc, but if pixar wants to render their next movie they use data-center compute.
> all those same algorithmic improvements would also be true for larger models
Smaller models run faster. If ten runs of a small model gets me the same quality result as one run of the big model, and the small model runs 10x faster, then they are functionally the same.
This is a very nice analogy actually and it impacts the whole story about US vs. Chinese leadership in "frontier AI".
It's always going to be cost;
developer time vs developer cost vs AI cost vs developer productivity.
With 4.6 it's looking like we are at the upper limit of appetite for cost (for "regular" Business) so the other levers will probably need to change.
It did ok, but scored substantially less than Opus. It also cost nearly as much, even with the current launch promo pricing for Deepseek.
That cost is interesting - I've seen similar things with Sonnet vs Opus, and in my own benchmarking there are some models that benchmark well, seem to have a good price but use so many tokens they cost just as much as "more expensive" models.
[1] https://blog.kilo.ai/p/we-tested-deepseek-v4-pro-and-flash
That depends on where the methodology goes. But more and more it's hands off. If the trajectory continues it won't matter because nobody is sitting their waiting / watching the LLM code anyway. It is all happening in the background. We might see hybrid approaches where the weaker / cheaper agent tries to solve it and just "asks for help" from the more expensive agent when it needs it etc.
My personal experience is that for production-grade code you need to steer the agent more often than not... so yes, at least some of us are watching the LLM code.
> We support the following backends:
> This project would not exist without llama.cpp and GGML, make sure to read the acknowledgements section, a big thank you to Georgi Gerganov and all the other contributors.Edit: aww, doesn't seem to support offloading to system RAM[0] (yet)
[0] https://github.com/antirez/ds4/issues/108
Guess I'll have to keep watching the llama.cpp issue[1]
[1] https://github.com/ggml-org/llama.cpp/issues/22319
Has anybody tried it? There is a lot of emphasis on MacBook Pro in this thread, but I would like to use it with an AMD Halo Strix with 128GB of unified RAM.
Configured one just now, delivers in 2 weeks
The code seems based on llama.cpp and GGML.
I don't fully understand why it is a standalone project. The readme discusses this: DwarfStar 4 is a small native inference engine specific for DeepSeek V4 Flash. It is intentionally narrow: ...
I think the only bigger difference in DeepSeek V4 vs other models is maybe the type of self-attention. And that leads to: KV cache is actually a first-class disk citizen.
But I still feel like those changes could have been implemented as part of some of the other local engines.
I also assume more models will come out, not just from DeepSeek but also from others, and they might share similar self-attention approaches, that would benefit from a similar KV cache implementation.
https://github.com/ggml-org/llama.cpp/blob/master/AGENTS.md
(the ux of ds4 is fantastic too -- it's dead-easy to get a known-good model, great quant. llamacpp you're much more hacking in the wilderness, w/ many many knobs.)
Is it true? We'll see, in a few years.
Antirez explained the dev process when he posted a pure C implementation of the Flux 2 Klein image gen model, at https://news.ycombinator.com/item?id=46670279
Right now I feel like a 4bit Qwen 3.6 27B with MTP is one of the best for agentic tool calling for some smart voice agents in an H200. I wonder if DS4 Flash being using 80b at 2 bit with 13b active and MTP could be even faster and smarter and allow more concurrent sequences?
This special 2bit quantization seems like a big deal.
It'll require some kind of:
- breakthrough in architecture or
- breakthrough in hardware or
- some breakthrough quantisization technique
The problem is that all the parameters need to be in memory, even the ones that aren't active (say for Mixture Of Expert Models) because switching parametrs in and out of ram is far too slow.
We show that EMO – a 1B-active, 14B-total-parameter (8-expert active, 128-expert total) MoE trained on 1 trillion tokens – supports selective expert use: for a given task or domain, we can use only a small subset of experts (just 12.5% of total experts) while retaining near full-model performance."
https://allenai.org/blog/emo
A crow exhibits some degree of intelligence in what is a very small brain compared to humans. There is overlap in the problem solving skills of the dumbest humans and the smartest crows.
So the question is: what is that? Yann LeCun seems to think it’s what we now call world models. World models predict behaviour as opposed to predicting structured data (like language.)
If your model can predict how some world works (how you define world largely depends on the size of your training data), then in theory it is able to reason about cause and effect.
If you can combine cause and effect reasoning with language, you might get something truly intelligent.
That’s where things seem to be going. Once we have a prototype of that system, there will be many questions about how much data you really need. We’ve seen how even shrinking LLMs with 1-bit quantization can lead to models that exhibit a fairly strong understanding of language.
I don’t think it’s unreasonable to expect to see some very intelligent low (relatively) memory AI systems in the next couple years.
prefill: 121.76 t/s, generation: 47.85 t/s
Main target seems to be Apple's Metal, so makes sense. Might be fun to see how fast one could make it go though :) The model seems really good too, even though it's in IQ2.
Can't say that it wouldn't be a better idea to spend that cash on tokens from the frontier hosted models though.
I'm an LLM nerd so running local models is worth it from a research perspective.
- More RAM: bigger models, more intelligence.
- More FLOPs: higher pre-fill (reading large files and long prompts before answering, the so-called "time to first token").
- More RAM bandwidth: higher token generation (speed of output).
So basically Macs (high RAM, okay bandwidth, lowish FLOPs) can run pretty intelligent models at an okay output speed but will take a long time to reply if you give them a lot of context (like code bases). Consumer GPUs have great speed and pre-fill time, but low RAM, so you need multiple if you want to run large intelligent models. Big boy GPUs like the RTX 6000 have everything (which is why they are so expensive).
There are some more nuances like the difference of Metal vs. CUDA, caching, parallelization etc., but the things above should hold true generally.
Wink wink, nudge nudge.
I have a feeling most cybersec researchers would only be interested in negative values of "reduce" :D
> Gentle reminder on how, in the recent DS4 fiesta, not just me but every other contributor found GPT 5.5 able to help immensely and Opus completely useless.
I've noticed the same for lower level squeezing-as-much-performance-as-possible code work.
I also don’t have time to do much personal coding outside of work, so I haven’t subscribed to a personal one yet. But I intend to go for Codex just to balance the Claude at work and also because of the hostile moves from Anthropic toward their consumer business.
https://gso-bench.github.io/
The long context reasoning is something I haven't even seen in frontier models - I was running at 124k tokens earlier and it was still just buzzing along with no issues or fatigue.
I am amazed at how well it works, I'm using it right now for some pretty complex frontend work, and it is much much faster than, for example running a dense 27b or 31b model (like qwen or gemma) for me (The benefits of MoE) - but the long context capabilities have been what have been absolutely flooring me.
Super excited about this project and hope Antirez can keep himself from burning out - i've been following the repo pretty closely and there are a ton of PR's flooding in and it seems like he's had to do a lot of filtering out of slop code.
Sure, MoE models have more knowledge, but extreme quantization may negate the benefits. And generally for coding tasks, you don't need a model that has memorized all the irrelevant trivia like, I don't know, the list of all villages in country X. DS4 also seems to run much slower on Mac Studio Ultra, which appears to be more or less in the same price range as RTX 5090. RTX 5090 gives me 50-60 tok/sec and 260k context with Unsloth's 5-bit quantization (only some layers are 5-bit too) and an 8-bit KV cache; prefill is instant too. It works flawlessly in OpenCode.
If you already have a spare high-end Mac, I can see the benefit, but I'm not sure it's a good configuration overall. Unless Qwen3.6 is more benchmaxxed than DS4 :)
Also have enjoyed playing with https://huggingface.co/HuggingFaceTB/nanowhale-100m-base (but early days for me understanding this space)
But I found its tool calling is reliable than other oss models I tried. I assume that it attributes to interleaved thinking. Its reasoning effort is adjusted automatically by queries. I enjoy reading these reasoning traces from open models because you can't see them from proprietary models.
I would love to try DS4 so bad. Well, I don't have a machine for it. I will just stick to openrouter. I wish I can run a competitive oss model on 32GB machine in 3 years.
You could try DS4 on that machine anyway and see how gracefully it degrades (assuming that it runs and doesn't just OOM immediately). Experimenting with 36GB/48GB/64GB would also be nice; they might be able to gain some compute throughput back by batching multiple sessions together (though obviously at the expense of speed for any single session).
It's so hard to predict what size the open-weight models will be, even in 6 months time. Will a 96GB machine turn out to be a complete waste of money? Who knows.
FYI, this to me points to an inference bug, bad sampling, or a non-native quant. OpenRouter is known to route requests to absolutely terrible, borked implementations. A model like DeepSeek V4 Flash shouldn't be making syntax errors like this.
> Starting from MacBooks with 96GB of RAM.
... oh. And I thought I bought a lot with 48 GB.
Here's one of the top hits: https://forums.developer.nvidia.com/t/fully-custom-cuda-nati...
Bizarre comment; sounds like "How do you know Porsches are fast? Did you drive one?"
I just find it really funny people are willing to write things like "empirically speaking, X is obvious" without actually testing it themselves.
I've seen mixed reviews, and the most honest sounding ones have said it has latency issues.
I don't really care that much what the average LLM power user says at this point, they're impressed by anything an LLM does. They're like toddler's entertained by the sound their Velcro shoes make.
You LLM people are going to be like my mom, once she got an Maps app she completely gave up on navigating anywhere with her own brain, and is lost without a phone.
Except for you LLM people, its going to be reading, writing, problem solving and thinking in general. You'll be completely reliant on an llm to get anything done, have fun with that. You're cooked bro.
"You LLM people". Has it occurred to you that individuals have variation within groups?
Apologies. Where did I form my opinions?