Kimi-K3 on HuggingFace(huggingface.co) Related: Kimi-K3 Technical Report [pdf] - https://news.ycombinator.com/item?id=49070985 |
Kimi-K3 on HuggingFace(huggingface.co) Related: Kimi-K3 Technical Report [pdf] - https://news.ycombinator.com/item?id=49070985 |
It’s a new fundamental innovation in math and CS that allows large scale lossy compression of natural language another data formats in a way that is semantically queryable and cross-referenceable. It also manifests some form of emergent intelligence, likely evidence of the long posited link between intelligence and data compression.
The tech is awesome. It’s one of the coolest things I’ve seen in over a decade. The industry is kind of shitty, which is not unusual. The discourse around it is almost universally insane, crazy people arguing with crazy people.
Oh and get off the AI eco bullshit train. Look up the energy cost of AI queries vs driving or running a home air conditioning system. Feel bad about using AI? Skip that DoorDash order. You probably just saved the energy of 1-2 days of heavy Claude Code use.
> If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.
I wouldn't be surprised if they started offering potentiall bad quantizations with much reduced capability at lower prices (without telling the users, of course)
Now the US government has 5 hours left to (attempt to) stop the release. (and save Anthropic)
Let competition run its course and the market (not government) determine the winners and losers.
On the other note, can't wait to see 1bit quantisation soon and how it performs in benchmarks, if it performs really well in benchmarks, would be very good news for GPU hosting providers, to offer "Opus 4.5 level model at the cost of Haiku 4.5"
Just checking in because y'all sound good with that.
Other than that they track Tool call error rate and Structured output error rate.
I only discovered this today, and it seems like a good idea. What are the problems in practice?
Getting 404 Sorry, we can't find the page you are looking for.
The separator is often the only way to distinguish American notation from ISO, so please use a dash for dd-mm-yy and a forward slash for mm/dd/yy
It's not practical to use it though. It counts almost 8x more towards your quota than glm 5.2!
Linux distributions got this right: Offer both HTTP and Torrents. Let the user decide.
Like they can takedown or update downloads and they can prevent someone from trivially bypassing the license agreements you need to accept for some models.
Perhaps that's exactly what they're trying to avoid, giving the user any form of control and having them depend on HF.
Perhaps let's call it Kimi-K3-2.8T-A104B to make matters clear.
If you have really high end hardware, you might be able to squeeze a heavily quantized version of Kimi-K3 onto your rig, but it will be too slow or too lobotomized to be useful.
This does put a near state-of-the-art open weights model within reach of what a small or medium business could afford if there's a case for local inference. It's probably not as good as Claude Fable or ChatGPT Sol. But if you're an organization that has a genuine need to run inference locally, this is a real possibility.
Is this for your homelab? Not in any practical sense.
Is this a possibility for organizations that can justify $1M or so on hardware for a near SOTA model they have full control over? Yeah, absolutely.
There are quite some countries around the world using d/m/y
https://en.wikipedia.org/wiki/List_of_date_formats_by_countr...
Algeria, Belgium, Brazil, Chile...
Also interesting to see what effort it will take to fine-tune this beast. The latest AISI benchmarks on cybersec place it above glm5.2, but still way way behind SotA closed models. Some fine-tuning might be needed here. Also, interesting to see if Cursor does another training round on it, to directly compare it w/ kimi2.6/2.7 fine-tunes (composer series) and grok4.5.
Also also, interesting to see if someone takes on distilling (proper distillation, w/ training the entire distribution) from this into smaller models. (dsv4-kimi should be really good, since dsv4 is very cheap to serve)
Huge price difference in what you can do with buying a used 4U rackmount server and putting 3TB of RAM in it (64GB DIMMs x quantity 32 in a quad socket xeon, you can see some benchmark prices on eBay for sets of 16 or 32 matched 64GB ECC DIMMs) for <$30,000, vs the cost of trying to run it on real GPU hardware.
Now obviously, as of the time I write this, the full precision hasn't been released nor has anyone like unsloth run it through quantization yet to produce a "Q8" or "Q8-XL" variant of it. But I think it's going to need more than 1536GB of RAM, with a usable and large amount of context, more like 2TB and preferably 2.5 to 3TB.
I also predict that people who try to run it in Q4 and Q6 will get the worst of both worlds, less precision/lost knowledge but also not reliable output that comes out too slow. In my personal opinion if I'm going to deal with something that is smart but slow and running on limited budget hardware, I need it to be Q8.
You'll spend ~100x more on electricity than the API cost to have it run on someone else's GPU at several hundred tokens per second.
I think some sort of extreme data privacy requirement is the only situation that justifies this, but the intersection of {needs absolute data privacy, needs to run SOTA model, cannot afford GPUs} is really really narrow. I wouldn't be surprised if this is an empty set.
Don't get me wrong, slow is sometimes better than "not at all", but depending on the performance, it might end up way too slow to even work for batched/async jobs like that.
So full CPU local AI inference may become viable option in coming years.
Or from SSD using something like Colibri[1]. Not going to be quick, but at least runable.
The model is known to be MXFP4 according to Kimi's release blog post, so the model weights will be less than 1536GB: https://www.kimi.com/blog/kimi-k3
Also, their previous models were native INT4, so it would be weird if they went larger now.
On a 3T model I’d imagine you’d be closer to 0.05 tks
I've made some proof of concept in https://github.com/woct0rdho/transformers5-qwen3.5-recipe . We can finetune Qwen3.5-35B-A3B in 16 GiB VRAM, and DeepSeek-V4-Flash (284B-A13B) in 90 GiB VRAM, without CPU offload. This works well on unified memory machines like Strix Halo.
Even so, larger models like Kimi-K3 still require multiple GPUs and nodes, and there are a lot more to do compare to single-GPU training.
It's almost certainly full parameter post training of the original model weights.
No, you don't. Without training cost you can infer only the marginal cost of serving this kind of models.
Moreover, you don't know the actual size of closed models (what if Fable is a 10T model? What if it's 1T?)
Still useful; "are the labs marginally profitable just on the marginal inference costs?" is still a useful question to answer. After all, if they aren't even profitable on inference in isolation, then we can expect to see large price increases.
If they are able to turn a marginal profit on inference alone, then perhaps the price increases won't be so severe (or perhaps they expand the time between generations so that they spend less on training but take longer to complete training).
"Are the labs profitable at all?" is, of course, a much more useful question, but that doesn't mean that the first question is completely useless.
Which is by far the most interesting number of the two.
> Moreover, you don't know the actual size of closed models (what if Fable is a 10T model? What if it's 1T?)
If you get close in output quality, then does that matter?
Of course inference efficiency is dictated by model architecture, size, etc. You can still guesstimate some of those and have an idea about cost/serve at several size tiers.
Are you talking about Kimi's training cost or the training cost of the model(s) that Kimi distilled?
Because Moonshot didn't even incur the majority of the training costs either
If you're going to open source your model, why would you set your own price high enough that other providers could easily and profitably undercut you?
well, exactly.
that's tough to answer without just average sampling because some users will ask the model "what's todays date" or "what color is the sky?" and some users will ask "Let's rewrite the linux kernel in brainfuck."
The Kimi-K2.6 model is 1.1T parameters with 32B active parameters. With light quantization (Q6_K) that's enough to run it (slowly) on a single 5090. On a single B200 you can have 5-6 experts loaded into VRAM at a time. Realistically that would be 3-4 to account for the context. [!]
[!] With this and other MoE models it looks like an interesting area for research would be to detect or predict which models would be needed ahead of time. That way you could schedule the load into VRAM step before the weights are needed. That way you shouldn't lose much/any performance from offloading the weights to RAM.
> With light quantization (Q6_K) that's enough to run it (slowly) on a single 5090. Kimi K2.6 is released as INT4 already.
So 5090 with K2.6 is just gonna sit idle 99% of the time, waiting for next slice of weights to load.
5.6 Sol calculates that single 5090 in raw compute & memory bandwidth can run K2.6 at 35 t/s (256k context depth) -- if it somehow had enough memory to hold whole model in VRAM. Man, I hope HBF succeeds and Nvidia brings it to consumer cards in 5 years..
Without leveraging system RAM and/or SSDs, I don't think you can, or how exactly are you running this, if this is something you are doing today? With CPU/expert offloading you could probably do it with a 5090 + 1TB of RAM or something like that, but absolutely not on a single 5090 entirely within VRAM.
Because Kimi K2.6 in Q4 is about 584GB GGUF size on disk and will use slightly more than that in RAM, Q8 is 595GB.
But my comment on running it was more towards serving this profitably at scale. You get much better throughput / unit of compute if you load everything in VRAM and serve many requests at the same time. That's how all inference providers are doing it.
Take a look at GLM 5 vs GLM 5.2 pricing -- GLM 5.2 cost more despite being the same model.
Take a look a look at DeepSeek, which hosts DS v4, profitably, yet others aren't able or willing to match the price.
OTOH, the multiple providers who have settled around the same price point ($3.48/M output tokens for multiple providers with good reputations) does indicate where it is profitable: https://openrouter.ai/deepseek/deepseek-v4-pro#providers
But I do think that the median price where this settles will tell us something about the floor at which it is profitable to serve this model.
> DeepSeek, which hosts DS v4, profitably
I specifically mentioned 3rd party providers, because there can be an argument that model creators themselves are subsidising tokens to gather training data for the next model. In fact, ds are public about their gathering of data (at least on openrouter they're marked as such). So that 0.x price point for dsv4-pro is likely subsidised.
they noted in their blog post they didn't focus purely on coding for grok 4.5.
> reduces it down to its minimum entropy -- it cannot be compressed further.
I think you could make a lot more money elsewhere :-)
https://en.wikipedia.org/wiki/Kolmogorov_complexity#Formal_p...
LOL
I think this release is actually great both ways when you think about it. We gonna be able to learn knowledge that labs have been hiding from us (e.g. cost like you mentioned). And labs could learn from whatever optimization techniques people come up with when trying to host this model.
It's honestly just good for everyone in my opinion.
Sounds like I'm buying a lottery ticket this week so I can drop $800k on hardware.
I am curios what's the most profitable thing to "plant" (agriculture analogy) on the land (cards) that you have have: web hosting, vps, llms, image/video generation, etc
> SemiAnalysis estimates that Anthropic's current blended gross margin has risen to the mid-60% range, with the API business gross margin exceeding 80%
Of course, people will insist "they are lying", "why should we believe them, it's well known they subsidize API pricing", ...
https://newsletter.semianalysis.com/p/anthropic-3q26-profit-...
https://finance.biggo.com/news/02d45650-b569-4d12-b44d-8d6d8...
Sota closed models don't even answer cybersecurity questions lol.
my friends whove tried in their companies gave up on it.
Any startup can download the weights, tinker with them, and fine-tune them. The real win here isn't necessarily cost, but performance on your data and IP sovereignty. It's a huge win. Kudos to the Kimi team.
Currently it's showing significantly better latency, but at a fraction of the usage Moonshot is experiencing, so we'll see how that holds up - regardless, a same-day deployment is an impressive feat!
I'm definitely having full eval suite on already if they get overloaded later on.
If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.
There is also this paragraph in their licence that is smart marketing-wise:
> 3. If the Software (or any derivative works thereof) is used for any of the Licensee's commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, "Kimi K3" must be prominently displayed on the user interface of such product or service.
It's either having a model struggling along with like 5-10 tokens per second on unified memory, or data center cards with hundreds of GB of VRAM consuming more than a kW of power. It doesn't seem like there's prosumer GPUs with like 180W-250W TDP and 128 GB or 256 GB of VRAM (one can dream). Then bifurcation and even just two of those cards would be kinda useful (albeit NVLink or equivalent would need to be commonplace).
Obviously nobody is running Kimi K3 locally without an insanely beefy homelab and lots of money to burn, but running GLM 5.2 would be cool at like ~100 tokens per second for a single session and maybe ~60 tokens per second with N subagents.
How unfortunate.
I saw arguments like "Providers cannot price less than their costs" in other comments. In economics, it's generally admitted that they shouldn't price less than their marginal costs, i.e. in their case roughly the cost of electricity, since a lot of these datacenters are not at capacity in terms of graphics cards usage (speculation since it's very easy to rent a GC for a couple hours on some providers). My guess is that someone will be selling tokens at less than electricity + depreciation of GCs soon, since there's a lot of competition and "smaller" data centers have overcapacity? This is speculation, correct me if I'm wrong
Over time the enormous investment in techniques and hardware manufacturing will almost certainly make these runnable in a more practical way. It will be a shame if by the time we get there it’s illegal to distribute them and you have to pay a reg capture premium and feed the machine.
In my testing it seems like Kimi has a healthy margin (I would wager 40-50% if they are renting GPUs at full marked up prices, a bunch more otherwise, given their tps, but I don't know which GPUs they are on and what they consider margins and if they own them) but definitely not the claimed 90%+ margins of Anthropic (honestly I am suspicious of even 80% API margins for Anthropic) as I have seen some people posit. If it was just electricity costs I could bet it could be 80-90% though otherwise it seems rough given the TPS they offer.
I would love if someone has access to those super secret R100s could try it, and tell us if it's significantly cheaper since I think immediate memory optimizations seem hard since I am already on a quantised model. And not even using 1M context.
All I had access to was B200(couldn't find a B300). I am certain people could optimize it a lot better but Kimi also wants some kind of contract for big providers so I think we shouldn't imagine any significant discounts while Kimi is the top open model around.
FWIW, China is suddenly talking about model export controls. It was one thing to release also-ran models, but now that they're pushing SOTA it's a different game.
Thanks for the hint. It makes sense, even if the main driver for releasing the models is to reduce the market size for the US companies.
Do you have any sources?
Edit: "market size"
It's just launch day gremlins, like always.
I think it's very valuable to have a smaller model that doesn't have any domain knowledge or facts built into its weights, but given the right context, could accurately reason about what to do and use the right tools.
I'm aware of colibri [1], but so far I've only seen extremely slow performance.
We won't be able to run this ourselves, but many providers can.
Hmm, not quite true, I think that honor, for better or worse, goes to OpenAI. When they released GPT2 (or GPT1 for that matter) is was quite literally the SOTA in the ecosystem when it was released.
is opus 5 a flop like 4.8 ?
where is the thread btw curios
You can still try it at https://chatjimmy.ai/, but it's running the rather outdated Llama 3.1 8B
The alternative is on-TPU flash for storing the weights.
Anyone know / is this a typo?
Even the python code for inference seems to use normal activations.
$3.00 / 1M In; $15.00 / 1M Out; 120 Tok/s
But since for reasons only Nebius knows Token Factory in general does not seam to offer prompt caching (at least not discounted) its essentially useless.
Raw fucking intelligence at your disposal, free to download.
If you'd describe what's happening now to someone from five years ago they'd think you're hallucinating or mad.
edit: the results I have seen from people trying colibri with fast consumer grade PCI-E 4.0 NVME SSD are 0.1 tok/s on models that are <700B in size, things that are well under 800GB on disk. With something that's 3T in size it'll probably be a lot slower than hat.
But still fun you can run it at home.
* Sparse Experts: 1481.4 GB
* Dense Experts: 1.9 GB
* Self-Attention: 72.4 GB
* LLM Head: 2.4 GB
* Embeddings: 2.4 GB
* Vision Encoder: 0.35 GB (surprisingly small)
plus some miscellaneous parameters.
Most importantly, we now know that the model has 104B active parameters, which is quite a lot and will make it difficult to self-host efficiently.
And then that's just for single prompts, what about agent harnesses, where before every tool call the model could reason a bunch?
I agree with you that real world results would be interesting, but I wouldn't hold my breath nor expect it to realistically be able to be useful. Still, people should try it, for science if nothing else :)
GGUF is maintained by all the llama.cpp developers. There are many quantization formats and algorithms under this container format, some are optimized for MoE (such as APEX quant), some for CPU and some for GPU, some work surprisingly well below 4-bit (and even near 1-bit). It also supports recent architectures like linear attentions and mHC.
The K3 maths can turn true only if the models size is roughly the same and the labs didn't find any better way to run inference at scale.
We don't really know that, for OpenAI and Anthropic. We suspect that, but as far as I know, even they have stopped claiming that they are profitable on inference.
So which is it?
Anthropic was probably profitable last quarter, without training costs: https://www.wsj.com/tech/ai/mind-blowing-growth-is-about-to-...
I other words, the providers that will be offering K3 inference don't have any training costs to offset, so they are only charging for the inference itself. OAI/Anthropic would need to offset their R&D and training costs in order to not be selling API access at a loss.
The distillation process involves getting conversation traces from the model you are distilling from, and then training your model against them.
You still have to do the training!
I guess this is one of the reason Anthropic i so "active" for asking for a development break.
It's worse than that: a typical MoE model routes a separate set of experts at every layer, not just every token! But in practice, RAM offload (for systems with non-unified VRAM) and even SSD offload still work surprisingly well given some amount of caching.
You can likely recover compute intensity and throughput by batching requests together, which (in practice, depending on sparsity) will end up reusing some of the loaded experts with high probability; though the obvious tradeoff is that having to store KV caches for the wider batches may leave you with less room to cache experts across layers and tokens.
(Plus if you're batching so widely that you end up loading essentially entire model layers, MTP then becomes applicable even for a MoE model. But this typically only applies if you're doing inference on a very large scale, or if your memory bandwidth is so scarce that you have to recover compute intensity by any means feasible.)
Caching really has nothing to do with this. With RAM offload you can mostly benefit from:
1) Batching for prefill is a huge win, even with MoE, since the batch sizes can be so large.
2) Keeping non-expert weights in VRAM, so the percentage of weights uses per token in VRAM is higher. This benefit reduces with larger models though.
> You can likely recover compute intensity and throughput by batching requests together, which (in practice, depending on sparsity) will end up reusing some of the loaded experts with high probability;
With MoE it's low probability.
> MTP then becomes applicable even for a MoE model
With MTP it becomes _extremely_ low probability.
There are a lot of optimisations that are not in the public sphere, source working on start up in this space
Sure, but if we're participating in public discussions, isn't it more fun if we talk about things people can actually read and understand, rather than secret stuff other's can say work, but no can actually validate or know how it works?
It sounds like "hybrid approaches are much better than the public is aware, because everything else is private and secret", but also: ok, so what? No one can run that anyways, (yet?), so why it matters?
> The Kimi-K2.6 model is 1.1T parameters with 32B active parameters. With light quantization (Q6_K) that's enough to run it (slowly) on a single 5090.
Is painting a very different perspective, even considering the latter parts it's hard to read that as "Of course offloading everything else that doesn't fit on the GPU itself". But anyways, it's been clarified now so no harm :)
For my product, I run GLM 5.2 and other models myself, in production, on rented hardware. Paying API prices would cost much more.
EDIT: You can now see several other third-party providers for Kimi K3 (Nebius, Fireworks). All charge exactly the same as the first-party. Does that mean that their costs are the same? Seems quite unlikely. It's simply not an efficient market, yet.
It depends on the total entropy of the model. Smaller models have less entropy.
Interesting. Why is that? I would have expected the opposite, since larger models have to try less hard to fit the training data. Or maybe this leaves more parameters with random initialization, resulting in higher entropy for larger models?
But if they're hoping to recoup non-recurring engineering costs rather than just hardware costs, they do need to consider the useful lifetime of the specific model.
Everyone keeps thinking those A100s only have 6 more months of life, and yet they're still going for more than they did per hour in 2024.
Show me evidence that A100 prices have collapsed, and maybe GPU depreciation will be relevant to the market.
However you should take it with a pinch of salt because IMO their direct quotations do not support their title.
Copy and paste below from my notes and reported memory consumption with latest llama-server, assuming use of "--no-mmap" to load the entire thing into RAM at the time that llama-server launches.
DeepSeek-V4-Flash-UD-Q4_K_XL via unsloth 145GB on disk GGUF 0.03.323.204 I common_params_fit_impl: projected to use 178175 MiB of host memory
DeepSeek-V4-Flash-UD-Q8_K_XL via unsloth 151GB on disk GGUF 0.02.215.885 I common_params_fit_impl: projected to use 184636 MiB of host memory
Laguna-S-2.1-UD-Q8_K_X via unsloth 120GB on disk 0.01.616.119 I common_params_fit_impl: projected to use 172860 MiB of host memory
Qwen3.5-122B-A10B-UD-Q8_K_XL via unsloth 160GB on disk GGUF 165GB RAM use on launch, fresh context 0.04.976.905 I common_params_fit_impl: projected to use 170038 MiB of host memory
But not every framework implements it properly yet.
Now whether that's good enough for one's use-case remains to be determined. You can also get more out of those (local models and quantizations) if you further tweak the harness you use them with, but tbh this is where it gets too much work (at least for me and the time I have available).
The "democratisation" talk from the frontier labs is especially egregious when they only release closed models (gpt-oss hardly counts) and are trying everything they can to make it harder to run open models.
I ended up implementing DiffusionGemma myself with Candle in Rust + CUDA, and it's quite literally the fastest model I've managed to run on my hardware.
Like, yeah if I could spend a few grand on such a GPU I probably would coz I'm a rich nerd, but I'd acknowledge it as an extremely inefficient luxury, kinda like a sports car.
So I think you could say the real misfortune is that we don't really have the technology (be it computer tech or political/social tech) to do that shared-HW thing in way we can truly trust.
Of course, by then we’ll want to run something commensurately larger.
128GB is enough to run a large model, quantized, REAPed, with MoE and fast SSD for model weights
Your power consumption estimates are off for this generation of GPUs. A 27B dense model gets 50-80 tps on an RTX 6000 using 600 watts.
1) Local LLMs are a relatively new phenomenon and hardware takes years. Apple probably lucked into their unified memory architecture being suitable (in terms of memory size and bandwidth) for local LLMs, but it's only with the newest generations we're hearing about LLMs even being a consideration in their design process.
2) NVidia seem to be deliberately blocking consumers from taking this path - as evidenced by the removal of NVLink from the 30x0 series onwards - probably to protect their data center cards from internal competition?
3) Perhaps there's just not the market for it? It's feasible that the number of nerds interested local LLMs is very small in numbers, sales, and profit potential compared to gamers on the one side, and data centers on the other. (This would explain why AMD and Intel aren't trying to out-innovate NVidia in this area, despite it being an obvious opportunity.)
Try to imagine output token speeds of 15,000 tok/s and a time-to-first-token of 200ms. (This has already been done for Llama 8B.)
Now imagine gargantuan context windows (2M, 4M, or even bigger); keep in mind the 1M context windows were science fiction a few years ago... now imagine having this on a local model on something like a phone or portable device that can be gathering data about things you're doing and constantly run inference for things useful to you. An obvious example of this would be a chatbot you can talk to that responds like a normal human conversation and doesn't have delays, but that's just scratching the surface.
So for prefill -- which is more about compute than bandwidth -- the Spark performs quite admirably. But on decode it's highly bandwidth constrained. Some smaller MoE models (e.g. Gemma4) can do 60-70 tok/second but anything dense, or anything that is actually filling up most of that 128GB is going to choke out around 15 tok/sec. Even at NVFP4.
My experiments with this are at https://github.com/rdaum/eider/
For my current work I get to log into trays on a real GB300. It's somewhat comical that NVIDIA is marketing the little baby on my shelf here as even in the same universe as that. Which is basically like having access to a super computer.
It will also happily hallucinate new names and content to fill in gaps in its knowledge, and present the hallucations in what looks like a correctly formatted sentence, so it could fool a person who doesn't know the subject matter. Like, I asked it for a description of Seattle and it hallucinated a name and description of a nonexistant tallest building in the city and suggested the view from its observation deck .
https://prismml.com/news/bonsai-27b
https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf
In no way was I surprised, it's asking a lot of under 6GB RAM usage. But I think for 99% of people they will get better results doing something over the network where the weights and inference engine are not on the device.
My guess is they are selling you the tokens, then selling your tokens (data) onto someone else.
The problems with frontier models (design taste, ability to solve novel/difficult problems, etc) cannot be solved by throwing more slop from the average user at it.
Actually, most of the main deficiencies in current models stem from the fact that their data sets aren’t curated and specialized enough.
Then there's 43% off at StreamLake with FP8 precision and 1M context window.
You are describing fundamental tradeoffs. Getting more performance relative to model size and training token amount is what all of the labs are solving.
Asking for such a model while wanting to be small and fast would align with what you're describing, which I believe is different from what I'm pointing to.
The model I'm describing sacrifices domain knowledge and expertise for agentic reasoning and tool-calling capabilities at a reasonable speed.
Think of Cactus Compute's Needle [1].
I’ve had success adapting even a 7B model for single-domain tasks that way, including reasoning and tool calling.
You can use an open model. The point is just to outsource the inference, so you don’t have to deal with running the larger model yourself.
Take the “train” portion of tasks on some benchmark, have K3 complete it, and then output detailed descriptions of tools used and why, then run the validation tasks with some small model that has access to the skills.
https://github.com/p-e-w/heretic
As a side note on this, if you see the reference in the screenshot in the link above to the harmful behaviors prompt set, these are all in English:
https://huggingface.co/datasets/mlabonne/harmful_behaviors
You could likely further de-censor a model by having a set of 'test' prompts in native Mandarin, Cantonese or really just about any other language. I don't speak any Chinese languages so I don't know if the published 'heretic' GGUF files some people have been throwing around will cooperate, or refuse, if you ask it in Mandarin for how to build a meth lab or precursors for semtex.
It is more important to focus on the questions than the persons who created it.
https://x.com/dhh/status/2081435006770249831 (from the creator of Ruby on Rails).
In his specific case Kimi did the task it was asked to do (translation of the article DHH wrote), which Claude refused to.
'When gypsies appropriate public spaces, you deport them. It's not hard, it's not cruel. It's the basic logic of self-protection.'
> The company appears burdened by far too many disparate groups that are over-optimizing for certain metrics as opposed to delivering usable technology for the company as a whole.
> And because Meta has a reputation for throwing money at problems and executing at high speed, these U-turns end up becoming more costly versus other companies that take a more disciplined or conservative approach. Suppliers also lose faith when given design wins are later cancelled. This has lead to less supply chain prioritization on new designs. Some suppliers favor focusing on Amazon or Google designs due to Meta’s frequent reshuffling.
> Few inside Meta’s chip division have a full understanding of why the company bought Rivos in the first place, and those who championed the deal internally have since gone quiet.
etc etc
It goes into a lot of depth.
https://newsletter.semianalysis.com/p/metas-infrastructure-t...
https://newsletter.semianalysis.com/p/the-future-of-meta-sup...
https://finance.yahoo.com/sectors/technology/articles/mark-z...
Not exactly the best use of dollars and the labor hours of some of the best minds of our generation.
Morale and momentum are huge things in companies, Zuck has been murdering both of those in Meta since... well naming it Meta.
If wonder if you can train a model to optimize this, by trying to make the expert selection sticky across a few tokens, without too much quality loss.
Another fun idea might be to try to build a model where the router chooses the expert 1-3 tokens in advance.
But the memory bus speed is fully committed when generating tokens or thinking.
0.07.015.888 I common_memory_breakdown_print: | - Host | 170038 = 162913 + 6740 + 384 |
The 6740 is the cache size.
There are parts of states like Grant County Washington that have cheap hydro power, but it's very rare for power to be that cheap in the US. Even if this applies to you, it won't apply to the vast majority of people on here who will have electric rates 2-4x higher.
Average electric rates by region:
New England 28.1 cents
Mid Atlantic 25.1 cents
East North Central 20.8 cents
West North Central 14.8 cents
South Atlantic 16.1 cents
East South Central 15.5 cents
Mountain 14.6 cents
Pacific Contiguous 26.1 cents
Pacific Noncontiguous 42.1 cents
Do you actually need to run the state of art model at 5 tokens per second instead of a qwen or whatever 7b or 30b model at 100 tokens per second?
Some people like doing things they want to do. Do I actually need to buy expensive pigments from europe to make paintings of flowers? My camera produces a much more accurate representation.
On the other hand, would it be cool to also have a really big thing as an ancillary tool that I could throw a request into opencode before going to bed, let it crank away and take a look at what it's done 7 hours later? Yeah, particularly if I (very much an unknown quantity at this time) could be confident that it builds high quality, syntax valid, appropriately commented and not absurd code.
Are there? At the highest levels of defense and law, AWS and Azure are used.
Having tried selling some of these entities on doing things in-house, there seems to be little interest.
This is certainly true if the user is an American company. You could look at the European initiatives to run this stuff on hardware they own in facilities they own and control within the borders of Europe for a counter-example.
Such as: https://www.google.com/search?client=firefox-b-d&q=schwarz+s...
https://www.dutchnews.nl/2026/04/government-turns-to-german-...
a) Contracting with a third-party independent inference provider who will run your choice of model on fast hardware that they own, with all appropriate data security/privacy/contractual/compliance protection in place
or
b) Contracting with the original creators of the model to run inference via their API and with assurances that all the same data protection is in place
or
c) Spending the money to buy your own inference hardware to run it on something you fully own/control at proper usable speeds?
Edit: Everything I've been writing in this thread is mostly within the context of being able to evaluate K3 and its usefulness to be self-hosted as a preliminary proof of concept or test of feasibility of a new thing, such as on <$20,000 of server hardware, before proceeding to spend 300-400k on GPU-related hardware, or external third party services/ongoing billing.
I have some inference I simply don't want to run on OAI, Anthropic, or Google because I don't want to run afoul of their "rules" and end up with a banned account, and this situation is only getting worse when it comes to doing fairly basic tasks like trying to secure your app against security problems.
A trillion-dollar business can easily trade dollars for the privacy. A business with $1M to spend won't even get a phone call with OpenAI or Anthropic, who were the only* previous players in town for doing this.
Worst-case example: Bootstrapped startup working in military.
It's also the case that an open model enables many more intermediate-cost solutions. E.g. providers certified for specific applications, on-prem rentals, etc.
* Omitting Azure, which gives some privacy for some $$$ on their models, but not at the level of high-security.
I don’t know if that’s 100x more than I’d pay (opex-wise) with an nvidia setup, but I can say the one-time capex is a great deal cheaper. Avoiding VRAM and DDR5 (fast DDR4 should be OK) are the biggest cost savers. ECC RAM is worth the extra price. General datacenter-quality hardware has less price sensitivity, and plenty of bang for your buck.
That's a world that I don't think we're ready for.
Young men 14-?? already compromise and attempt to extort organizations daily, sometimes cluelessly from western nations, often not. It doesn’t have to be gangs when the home country doesn’t care / isn’t technologically or culturally developed.
Already seeing AI-written payloads and frameworks in the wild. I think it’ll turn out that AI won’t build you a maintainable ERP but it can create C2 networks, exploit POCs or even 0-days potentially, and let kids make their own ransomware tooling. Then we’re dealing not with a handful of cybercrime tool makers but a generational problem.
If you want to assess the position of the U.S. Copyright Office for yourself, the relevant text can be found in the Compendium of U.S. Copyright Office Practices § 313.2, "Works That Lack Human Authorship"[2], which states:
> […] the Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a) (emphasis added). To qualify as a work of “authorship” a work must be created by a human being. See Burrow-Giles Lithographic Co., 111 U.S. at 58. Works that do not satisfy this requirement are not copyrightable.
> […] the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author. The crucial question is “whether the ‘work’ is basically one of human authorship, with the computer [or other device] merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.” U.S. COPYRIGHT OFFICE, REPORT TO THE LIBRARIAN OF CONGRESS BY THE REGISTER OF COPYRIGHTS 5 (1965).
Oh, and there's also a bit in the following Section 313.3, "Works That Do Not Constitute Copyrightable Subject Matter"[2], which explicitly excludes mathematical principles, formulas, algorithms, and equations, along with DNA sequences and other genetic or chemical compounds, regardless of whether they are produced by humans or by nature. If one takes the perspective that machine learning models are algorithms, the conclusions on copyrightability are pretty clear.
[1] https://media.cadc.uscourts.gov/opinions/docs/2025/03/23-523...
[2] https://www.copyright.gov/comp3/chap300/ch300-copyrightable-...
They'll give you HIPAA compliance, they even have a data center for US government classified data, they can give you European data sovereignty. And with OpenAI and Anthropic models to boot, you don't even have to settle for open weights.
What kind of privacy needs do you really have beyond that?
Even for EU companies running open weights on EU stacks LLM inference on the GPU must process plaintext and I can't find any EU provider with NVIDIA H100/H200/Blackwell CC mode plus SEV-SNP or TDX, where you can cryptographically verify the workload ran somewhere the operator cannot inspect.
Personal compute is therefore the only option if you want personal autonomy privacy for IP &c. Maybe another option is to use cloud compute rented to fine tune a personal model that suits your own needs that would help bring the cost down, I don't know enough about this area to know if it kills the "intelligence" of those domains due to limited ?cross-verification within the LLM.
That's the easiest case.
AWS Bedrock models running in AWS Secret Cloud for Industry. (I really have no affiliation with them, I'm just like... this is a completely solved problem, why do people think this is hard and requires on-prem hardware?)
https://www.aboutamazon.com/news/aws/aws-secret-cloud-for-in...
I'm with GP that these are tinfoil hat concerns, when there are solutions to all of these, unless you're perhaps in some country with very specific needs beyond things like European sovereignty or US military secrets (like a non-US defense concern).
Note that the other commenter never said US-based military oriented startup. You just assumed, then jumped to "heck yeah let's use Amazon Secret Cloud for Industry"
Not everyone has or wants an office in Crystal City.
If I were ranking third parties on their ability to safely handle my data without compromising it, I would rank Anthropic pretty low for things like Fable (where they more or less promise that they will misuse my data), but I want Azure pretty low in the sense that I fully expect them to be compromised.
I would tend to trust Amazon to avoid being compromised.
https://www.hydroquebec.com/residential/customer-space/rates...
Another example would be Manitoba hydro
All figures in Canadian currency
Under heavy inference load you will find that the cpu usage is actually less as the bottleneck is the RAM bus speed. An older 2U rack server that is 600W load (typically a 1+1 power supply server when plugged into two kill-a-watt would show 300W on each, equal load balancing) when maxed out with stress-ng might be only 450W total running inference.
If you have 600kWh used in a month by running something 24x7 and your power is $0.15 a kWh, that's more like $90/mo (not counting cooling or any ancillary costs for the environment where it's in).
The options for AC power supplies for servers with 2 or 4 load sharing redundant power supplies are a lot greater. If you were buying all new hardware and starting from a clean sheet of paper design with lots of money to spend, absolutely. At that point also start looking at higher voltage DC distribution stuff related to open compute platform and 800VDC.
But if I were trying to make the absolute most use of $20,000 to put together a 3TB RAM server (48 x 64GB DIMMs), it would likely end up AC powered.
Where in the world are you finding that much RAM in a racked server for $200/month?
https://www.macrumors.com/2026/06/25/m5-ultra-mac-studio-202...
I've seen public pronouncements that the RAM shortage could persist for a decade.
And then if consider that the constraint on local LLMs isn't just memory size but bandwidth ...
If you take something like a DGX Spark and increase its memory to 512GB that doesn't even solve the problem. Because the bandwidth of DDR5 just can't manage reasonable speeds for decode. If you take a dense model or an MoE model uses up most of that 128GB in active decode you will only get like 15 tok/sec. "Real" datacentre inference boxes use high bandwidth memory that is 10x the speed.
I think we're unfortunately a long way off, unless people learn to accept working with much less intelligent models locally.
The innovation is going to have to come on the research & software side -- we need to find ways to pack more intelligence into a smaller number of parameters.
Filtered heavily into those who care only about money. Many don’t want to work there. Smart people have other choices.
Smart? Check.
Money-obsessed and soulless? Check.
When you're trying to estimate/infer the costs of serving the tokens and even include the cost of training the weights in order to output tokens then yeah, why wouldn't that matter?
Only if you don't have to continuously train new models, and you are not at a runway risk.
> April 2016, 8 GB was standard across the 13-inch MacBook Air range
... Now it's 16.
Rich nerds will have quite a bit more. But I suspect the standard of model rich nerds want to use will have gone up somewhat too.
It's a 3T parameters model, with a weight format (MXFP4) still not completely integrated into the ecosystem, which only a few has the hardware to even do inference with, much less fine-tuning or more post-training. But sure, do sit and wait a few days :)
However, even if it did, it is absurd for an AI to decide what it will and won't translate.
You can!
> AFM 3 Core Advanced makes routing decisions per prompt. A lightweight, dense block selects a fixed set of experts during initial processing, periodically reselecting them during generation.
https://machinelearning.apple.com/research/introducing-third...
OpenAI and Anthropic guard the real reasoning closely. Users have never been able to see it and the API returns an encrypted blob instead of legible reasoning.
Here is an archived version of Anthropic's API docs saying that Sonnet 3.7 (only) has unredacted CoT on API: https://web.archive.org/web/20260324051339/https://platform....
5.5 Sol repors 10x fewer reasoning tokens than Kimi k3. If it is correct, than it unlikely has those doubt issues.
At the same time, I feel like their reporting is incorect and we are now paying per "intelligence", not actual tokens. We can't verify it anyway..
I think it's marketing, advertising, and product refinement.
Ex: all the things Google wants your search data for.
It's somewhat silly to think the value of that data has changed much. Advertisers want to know what's popular and getting clicks and attention. Competitors want to know what features are getting used in their markets.
In the simplest case, think of this data as improving the harness, not the model.
That would be valuable to advertizers for example.
How are you going to use tools, if you don't have the context to use them?
Imaging if you could only think in terms of lambda calculus, and was asked to check the weather to let me know if the weekend is good for a hike.
(If you're claiming that it is impossible, and you have references you can share, then I'm honestly interested.)
I think more of the capabilities comes from the cross domains knowledge.
Anyways, software is also designed for a domain. The reason why it is so adept at making a fitness tracker is likely because it knows about biology.
Just my experience though, I'm still figuring things out. Perhaps some subsets of tasks would be more ideal for these long-horizon workloads - exploration, multiple competing implementations, etc...
I left something gargantuan running over the weekend (decompiling 1980s-era system software) and look forward to checking it out later today when I have a few free minutes.
Given the hardware shortage in the world, I suspect renting ("sharing") via APIs will likely remain cheaper for the foreseeable future since each piece of hardware isn't sitting idle nearly as much.
That's not an emerging practice, it's a tested strategy that is these days only used as a last resort by those desperate to fit a model in memory. Some models do better than others, but generally the model quality suffers greatly under those conditions.
But it's a moot point, because for local inference on consumer hardware, the MoE is so much faster.
An alternate theory is that models need lots of input knowledge to learn complex reasoning, but don't need so much at inference time. An example is arithmetic. Early in their training, models do arithmetic (poorly) by pattern-matching on memorized examples. Eventually they grok arithmetic and stop pattern matching, and then they don't need to remember the examples anymore.
The fitness tracker doesn't take much knowledge of biology. Large models have quite a lot of biology knowledge that most people will never use. Same goes for lots of other topics. For basic knowledge that easily fits in context it could search the internet, or a local collection of introductory textbooks.