Recently I started playing with LM Studio and local models and found out that `gemma-4-26b-a4b` is suprisingly capable. I don't need elaborate akin to "create complete app to do X" or "refactor the whole codebase of bazzilion of LOC" but rather "how to go about doing thing X" or for language study (explaining nuances of phrasal verbs or subtelties of vocabulary in other languagues) and darn -- the results are rather good (to the point that I use it mostly nowadays)

>which box to buy

…

> spark, costs less than a conference trip.

I know putting actual prices regionally localizes your article and temporally, with how prices are so unstable. But analysis of “what to buy” without actual prices is borderline meaningless.

Overall, good article, very interesting to see a real deployment that’s actually attainable and not just a subscription to a big 3 token plan.

I could not focus on the article with all the noise in the background. It was annoying in the header but then it continued down the page. If you want to do this on your marketing pages have at it but for a blog/news style page? Reader mode was the only way to restore sanity.

I would love to see such comparisons but with quantized versions, because quantization allows running models on smaller hardware with some quality loss. I am running Qwen3.6-35B-A3B quantized to int4 on an A6000 card that was otherwise just sitting around idle. It works up to a degree, but I would love to see benchmarks comparing different quantizations of several models, especially in quality. This is an important dimension in deciding whether to buy a GPU is worth it, and it is missing from this (otherwise comprehensive) article.

This seems like a temporary situation. Utilization maximization is a matter of switching to the software factory unattended flow and intelligently routing tasks that will not require assistance to the unused time.

A decade and a half ago we used to run massive map reduce jobs overnight. Code will be handled like this.

Right, so the 1.4 TB weight footprint straight away makes me wonder how much more important load/unload time becomes in terms of the neocloud product here. With 1.4TB of weights, a fast-enough cold start would basically be _the_ product; fast enough weight-swapping that can make K3 into something you could usefully rent by the hour/workday.

Interesting because again, the license Kimi shipped under [0] defines "Model as a Service" as

> giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data

and then they have that clause around if you operate such a business above $20M aggregate revenue you need a separate agreement with Moonshot before commercial use, which presumably captures the majority of the larger neoclouds best placed to optimise this.

But where's the line? Say you offer infra optimised for GPU inference, warm pools, isolation per customer, exposed control plane, billed per GPU-hour rather than token, the invoice says compute rather than calls. The customer arguably 'self-hosts', you're probably fine? And if you as a provider run the serving stack and hand your customer an inference API, you're inside the definition regardless of whether you charge by the second or by the token. But what about if you give them direct hardware access, but have the weights cached on NVMe / ramdisk hyperlocal to the infra they're renting so that their hf cli pull only takes a few seconds? Sure, a managed warm pool of GPUs with K3 pre-loaded probably isn't ok, but a local hugging face lru cache holding 'whatever your customers pull down most often', superoptimised for fast weight swaps that the customer controls... is? Is it?

Again, where's the line? Is it materially different from a local docker registry mirror? What about safetensors checkpoints pre-sharded for the specific hardware topology you're offering? Does it matter whether you perform the checkpoint optimisation yourself and make it available, or merely cache one published on HF that happens to target exactly the hardware you rent out? What if you published that checkpoint yourself?

I'm definitely overthinking this, and I'm sure there's been conversation here about this already, but the other kimi threads[1] are enormous. And I am curious.

I'm also curious to know whether Moonshot would actually be against a setup like this. Guessing they would if it was AWS (not quite elastic but not that dissimilar), but what about others? Realistically I guess it'd be easier to just talk to them, especially if you were doing it in a way that targets a slice of the pie they never would have gotten anyway due to data residency requirements etc..

[0]: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE [1]:https://news.ycombinator.com/item?id=49065752

Why does the website have to be so trippy when scrolling? It fucks with my eyes.

Author here(I'm on the team). We updated this post after Monday's Kimi K3 weights release: fitting the 2.8T model means going from 8×B200 to 8×B300, ~20% more hardware cost, and concurrency drops from 24 to 16 users vs GLM-5.2. Caveat we're upfront about in the post: our 64-task SWEBench Pro subset may be in Kimi's training set, so the 86% resolve rate is an upper bound.

Let us know your thoughts, we really value feedback!