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.
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?
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.
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.
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.
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.
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.
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).
On a 3T model I’d imagine you’d be closer to 0.05 tks
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.
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.
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 :)
Or from SSD using something like Colibri[1]. Not going to be quick, but at least runable.
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.
But the memory bus speed is fully committed when generating tokens or thinking.
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.
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?
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.
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.
I guess this is one of the reason Anthropic i so "active" for asking for a development break.
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
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.
> 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...
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.
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.
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 :)
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.
> 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..
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.)
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.
they noted in their blog post they didn't focus purely on coding for grok 4.5.
> 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...
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.
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.
0.07.015.888 I common_memory_breakdown_print: | - Host | 170038 = 162913 + 6740 + 384 |
The 6740 is the cache size.
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).
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.
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.
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.
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.
128GB is enough to run a large model, quantized, REAPed, with MoE and fast SSD for model weights
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.)
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
My guess is they are selling you the tokens, then selling your tokens (data) onto someone else.
Then there's 43% off at StreamLake with FP8 precision and 1M context window.
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.
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'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 :)
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.
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.
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.
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.
signed: a hungarian :)
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
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...
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.
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.
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.
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.
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)
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?
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.
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)