Kimi-K3 on HuggingFace

It will be very interesting to see what kind of 'slow' performance people get from running it on a no GPU, but tons of RAM server (like a dual or quad socket xeon with 1.5 to 3TB of RAM). For the purpose of giving it longer duration tasks to generate a piece of something and come back and check on what it has done in 4 or 6 hours. Even if the output is like 5-6 tok/s, that might be usable for some purposes.

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.

    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

Speaking of finetune, currently a common practice is LoRA over bnb 4-bit base model, but I think it's time to replace bnb with GGUF as the base model format. GGUF is actively supporting new model architectures and more aggressive quantizations.

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.

AISI is capped at 100M tokens and K3 is less token efficient than Anthropic/OpenAI models. There is an argument to be made, looking at AISI results, that with uncapped tokens it would be just slightly behind the closed weight players.

> Then we'll be able to guesstimate if "labs are subsidising tokens on API pricing".

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?)

Since the model is natively MXFP4, I think it'll be even more interesting on the hardware front. It'll comfortably fit on a 8x AMD MI355X node. I suspect that'll drive token prices down, further.

Say a single Kimi K3 is deployed on 16 x B200s: how many concurrent users can that handle? I realize the question assumes a major simplification that everyone's prompts/sessions are the same.

You're assuming inference providers are going to sell tokens at cost. You're also assuming that the inference providers have will optimized inference engine. I haven't seen that to be the case so far, to be honest.

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.

If it is a mixture of experts (MoE) model like the 2.x models, won't this reduce the hardware needed to run the model?

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.

If it takes so much resource to run, how does the sharing of a single llm works? There is some interface that basically submits context/cache plus current promt, from each user, doing basically time-sharing compute?

We have a lossless compression codec (working on open sourcing it over the next couple of weeks) that reduces it down to its minimum entropy -- it cannot be compressed further. On all tested large models, it's a ratio of 1.34-1.23 -- and smaller models up to 3.76x. It also increases the effective bandwidth by the same rate.

This is a very insightful eye opening take. I haven't even thought of it this way. This really is the first open model to be as big as the frontier has been until now.

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.

>realistically you'll need 16x for context / throughput optimisation

Sounds like I'm buying a lottery ticket this week so I can drop $800k on hardware.

I think cursor will likely do a grok fine tune rather than a kimi one for the next composer.

they noted in their blog post they didn't focus purely on coding for grok 4.5.

So nowadays the hardware and hosting providers must be in an optimization race, whoever can make the model just a bit smaller or more efficient (to fit on fewer/less powerful cards) will have a huge advantage and can make a lot of money.

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

> The latest AISI benchmarks on cybersec place it above glm5.2, but still way way behind SotA closed models.

Sota closed models don't even answer cybersecurity questions lol.

> if "labs are subsidising tokens on API pricing"

> 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...

pardon my ignorance but is fine tuning still considered viable in face of rapid model releases. Is it really worth it?

my friends whove tried in their companies gave up on it.

Anyone who thinks that the labs are not profitable on per token API pricing is delusional and hilariously wrong.

[flagged]

Fireworks' priority tier of Kimi (at $3.75/M vs. Moonshot's $3.00/M) is available on OpenRouter as well. https://openrouter.ai/moonshotai/kimi-k3#providers

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!

Comparing to Opus 5: Claude Opus 5 (Uncached Input $5/M Cached Input $0.50/M Output $25/M) but you also pay a premium on Cache write 25% for 5m and 100% for 1h.

good find! This sounds a bit like what Meta was doing with the earlier Llama models?

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.

I have found that the "mostly didn't lose anything" Q8 large models that I want to run are all too large to run on the "only $3995!" 128GB max RAM systems that some people are buying, and definitely won't fit with any usable amount of context. Things like Qwen 3.5 122B Q8 or deepseek v4 flash Q8, or Laguna S 2.1 Q8 need 170-190GB of RAM including full context, which fits on a 256GB RAM dual socket workstation or rackmount server (sans GPU).

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

Post-crypto, the GPU manufacturers took the proactive move to use VRAM to segment the market for the purpose of price discrimination. Sure, data centers will pay vastly more for GPUs, but Nvidia knows that the PC market is steady and reliable. They could get the best of both worlds by kneecapping their consumer cards to tiny amounts of RAM, to dissuade the cloud providers from scooping up all the consumer cards, and then charging the two segments wildly different amounts for what amounts to the same hardware (back when the cost of RAM was negligible)

LLM inference unfortunately also seems to be a task that's poorly formed for moderate consumer hardware,as a single user. For a single user use case, the load is bursty but requires the weights to be in memory already. So a multi user server that keeps the model weights in parts of its memory and then spends some more per user kv cache is wildly more efficient and the wildly expensive gpu cores aren't just sitting idle most of the time. Don't get me wrong, most desktop workloads are bursty, but the power needed to to them has gotten cheap enough that we can have way overkill for idle scenarios hardware just sitting on our desks.

I've been feeling for a while that as we keep increasing model size, we're going through the opposite of the PC revolution.

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.

Yeah, I've been experimenting with DiffusionGemma which sadly isn't as smart as Gemma itself, but holy hell is it FAST, and has image input as well, so doing things like "take a screenshot once per second + ask the model to categorize/model it WITH reasoning before" becomes realistic and doable.

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.

I agree but worth noting that it's never gonna be very practical to run LLMs like this at home. Unless we have some sort of design breakthrough, the only "sensible" way to run them is at high batch levels on shared HW.

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.

The next mac Ultra will allow to run a big model locally with acceptable speed. But we need people to optimize it for that computer, and we’ll be more limited in models we can choose from

128GB is enough to run a large model, quantized, REAPed, with MoE and fast SSD for model weights

> like 180W-250W TDP > running GLM 5.2 would be cool at like ~100 tokens per second for a single session

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.

RTX Spark does go up to 128gb No idea how it compares tho

You're right, and it's interesting to consider why. It's probably a combination of a few factors:

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.)

The individual-shaped-hardware problem gets even sharper at the phone end. Shipping a 3B model on-device, the usable RAM budget after the OS and everything else is more like 2-4GBtotal, not per-model so it's not 'can I afford more VRAM', it's 'can I fit a language model and an STT model and embeddings without the OS killing my process'. Feels like phones are the most hardware-constrained 'individual' tier and get the least airtime in these kind of discussion. Is that because the models that fit are still too limited to be interesting, or something else?

Seminanlysis is estimating sub $1 cost per MT for ~2Trillion models. The numbers change based on throughput and quant, but it is conceivable that provider costs at scale are low enough that even $2.42 per MT on GLM 5.2 (current best price) is margin positive by a wide margin.

> 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

My guess is they are selling you the tokens, then selling your tokens (data) onto someone else.

Press x to doubt on the 45% number. The cheaper providers on open router are fp4 vs fp8 for official zai. There are some cheap fp8 ones (like novita) but the ui makes it seem like it's a temporary promotion, with their normal prices being almost equal to official zai (idk much about open router so not really sure what's going on with these discounts)

[dead]

Until a few minutes ago there was a countdown page. (The weights haven't been released yet.) The countdown should be over in 19min, not sure why we're suddenly getting a 404.

It's up now

I imagine it's just technical issues on the flip. It's also going to be interesting what happens to HF with loads of people downloading a many TB model. Even though almost no one has the capability to run it, it does seem like something to stash away in case it suddenly becomes unavailable due to government controls.

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.

At the risk of sounding like a conspiracy theorist, this sounds like a great opportunity to make a statement. US or China, but likelier to be the former. Maybe Clem's on a call with the US government right now?

There is a really interesting startup in Prague that is doing just that. They fine-tuned Qwen 3.6 27b to have 46% fewer reasoning tokens while maintaining most of the performance characteristics. I'm interested to see if they continue down this path of optimizing reasoning for other models.

https://bottlecapai.com/post/thinkingcap-qwen3-6-27b/

Yeah, they thing forever and doubt everything "wait but" for 200k tokens for almost any question.

There's a 2bit quant on HF already at ~1TB

"I'd like a car that goes 300mph and gets 100mpg while doing it. I'm aware of a car that gets 100mpg but it is extremely slow."

You are describing fundamental tradeoffs. Getting more performance relative to model size and training token amount is what all of the labs are solving.

Why not generate an artificial dataset using commercial APIs and then finetune a small model on this data?

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.

There’s another way besides distillation that’s way cheaper: You can have the big model build prescriptive skills that the small model follows.

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.

What grip do they have then ? Is Claude code and the suite of other interface that good compared to open source/ competitor offering ?

> (there's even a thread on HN about Opus 5's problems

is opus 5 a flop like 4.8 ?

where is the thread btw curios

> For the first time, an open-weights LLM is right at the top.

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.

As a completely one person, single sample anecdote, the 'heretic' uncensored Q8 GGUF variants several people have published of Qwen 3.5-122, 3.6-27B and 3.6-35B-A3B will very happily discuss just about any controversial topic that the CCP hates. Including lots of things that would get you thrown into prison if you published them in Mandarin on the domestic Chinese internet.

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.

Outside of asking it to talk about Tiananmen Square, are there any standard tests for "bias"? And if so, who created them and what are their biases?

It would, indeed, be interesting to compare, given what we know about Anthropic’s censorship and political bias in their closed and more expensive models.

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.

There's is always one in each and every AI forum: "But, but have you asked it about tiananmen"?

[dead]

The SemiAnalysis piece on this is long but very much worth reading:

> 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...

The cynic in me says maybe they would be further along if they hadn't spent $80 billion on trying to build the "Metaverse" VR world. I've never met anyone who actually uses it and to the best of my knowledge it has very low mass market uptake.

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.

Because lack of talent and organizational disfunction matters a lot more than you think. The reason why OAI and Ant are always at the top is because of this and I’d say compute is third on the list.

because imagine starting work every week and finding that your dumb ass CEO pivoted the company again and is ruining other peoples lives, and he then reorgs the management again so you now have your 5th leader this year.

Morale and momentum are huge things in companies, Zuck has been murdering both of those in Meta since... well naming it Meta.

Turns out sitting quietly in a room and doing math is worth more than all the money and network in the world.

At this point in time what does meta get from releasing open weight models? Why devote the resources to it.

The path to ubiquitous AI (17k tokens/sec) https://news.ycombinator.com/item?id=47086181

You can still try it at https://chatjimmy.ai/, but it's running the rather outdated Llama 3.1 8B

Yes, from 6 days ago: https://news.ycombinator.com/item?id=48986351

Given how fast models are improving, burning the weights into actual ROM is prohibitively expensive if you need (or want) to replace the chip every couple of months.

The alternative is on-TPU flash for storing the weights.

Yes. We're still a ways off from this being ubiquitous, but I am convinced it's coming.

If it's 4-bit native for the sparse parameters (which is the bulk of them) why would you expect it to add up to 2+TB?

It's up.

Unfortunately I don't think modern consumer CPUs are physically capable of addressing enough RAM to even load the model into memory. We'd have to wait for some random person to make an extremely quantized version before we could reach those blazing speeds

Coreweaves gonna boom

I think we felt the same when Apache or MySql was releasing back in ancient times.

China uses YYYY/MM/DD, which is logical.

20270727 if we're improving dates :)

or 27/7 for the rest of the world

Or 11/Shrawan 2083 if you're in Nepal

I just checked and I have it available in opencode go! I just tested it with one message and confirm it works.

It's not practical to use it though. It counts almost 8x more towards your quota than glm 5.2!

https://opencode.ai/docs/go/#usage-limits

[flagged]

Yeah, why not. Toughest part is running the hardware so you can create the traces for downstream training, but once over that hump, nothing would stop you from doing that no.

For huge models like these, the only reasonable way to host them is via torrents. I don't understand why hf doesn't offer this as an option.

Linux distributions got this right: Offer both HTTP and Torrents. Let the user decide.

I've had a steady 2 gigabits (my maxxed out ISP bandwidth) since I started the download at launch.

It’s going to get very little downloads just by virtue of size. Very few shops will be able to directly use it

The short answer is no, it won't work on your home computer. In it's current form it needs something like 594 GB of memory, far outside what you can reasonably run on normal consumer hardware in 2026.

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.

Yes it will. Buy a 4TB nvme, allocate 3TB as swap, and run a gpu emulator on your cpu.

Right

You're not having a party?

It saves you from having to perform date-time calculations.

It's a release party

I don’t need the fastest car to get to where I’m going. I need a car that gets me to where I’m going at the speed I’m comfortable driving at.

Modal eng here. Getting the model running is quite challenging, but its accessible right now on Modal via Endpoints: https://modal.com/blog/kimi-k3-by-moonshot-now-available-on-...

No, luckily private companies in the US are aching for this to NOT be banned. NVIDIA, Microsoft, etc. just released that letter. We’re saved from the trillionaire companies (OpenAI, Anthropic) by the other trillionaire companies acting in self-interest (hosting and hardware).

I don't know if the masses can quite afford the 500k in GPUs you need to run this

a "moment for humanity"? as if this shit isn't going to generate 99% slop at the cost of all we have left as a species?

They cant push it too low. The license agreement it is released under wont allow it.

> 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 think the results might be underwhelming - AI providers need to turn a profit and can't subsidize, and they're working off of the commodity hardware everyone does.

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)

As long as they are transparent about what quant they serve the model and any other optimization they do that also affects performance of inferred tokens.

I have a feeling the number of people downloading 2t parameter models is significantly smaller than the 1-100b models.

I think China publishing this stuff is more about prestige. The US has had export controls that make it illegal to sell Nvidia chips, and other AI hardware to China, and this is China saying "yeah, whatever". Also, it weakens western tech companies' position in AI, and pushes CCP bias perniciously. Building your product/company on top of a text-generation model that favours the CCP position on everything is just peak propaganda.

if it's anything like the speeds Nvidia Nim puts Deepseek at, it'll probably be 10 tok/s or lower and timeout frequently

Doesn’t it refer to the emoji?

Those are facehuggers.

haha stupid xenomorphs with their acid blood and pointy bits - all they had to do is make the beasts write code and do our homework :)

Funny retcon, but come on… TIL huggingface started as a chatbot for teens who didn’t get enuf hugs.