Kimi K3 is the largest open model ever — and it splits "open" from "runnable"

A frontier-class coding model is now something you can download — but K3 at 2.8 trillion parameters quietly pulls apart a word that used to travel as one idea. "Open weights" has become two separable things: a license that lets you own, inspect, fine-tune, and redeploy the model, and an infrastructure bill you have to pay before it does any work. For a 7B or even a 400B model those arrived together. At 3T-class scale they don't, and pretending they're the same is how a team ends up surprised by the GPU node it never scoped.

Moonshot AI released Kimi K3 on 17 July and scheduled the full weights for Hugging Face on 27 July under a Modified MIT license — the same permissive lineage as its earlier models. K3 is a 2.8-trillion-parameter sparse mixture-of-experts model, the largest open-weight model anyone has shipped, and it activates just 16 of 896 experts per token through what Moonshot calls a Stable LatentMoE framework. It carries a one-million-token context window and a new attention design, Kimi Delta Attention (KDA), that the lab reports decodes up to 6.3x faster at that length.

On benchmarks it is not a novelty act. K3 tops the Frontend Code Arena leaderboard at 1,679 Elo — ahead of Claude Fable 5 and GPT-5.6 — leads SWE Marathon and Program Bench, and lands within three points of Fable 5 on the Artificial Analysis Intelligence Index. In other words, a coding model roughly at the closed frontier is now something you can download and own outright.

The reflex is to file K3 next to every other open-weight release from a Chinese lab, and the coding scores invite exactly that. But the number that reframes the launch is the parameter count. Earlier open models you could plausibly run on one workstation or a single accelerator; K3 at 2.8T cannot be. Its weights run to hundreds of gigabytes even quantized to four bits, and serving it in production means a node on the order of eight H100-class GPUs with an inference stack — vLLM, SGLang, or TensorRT-LLM — that understands 896-expert routing.

That is not a footnote. Moonshot shipped a vLLM path alongside the weights precisely because standard serving does not handle this out of the box. "Open" here is completely real and completely different from "runs on the hardware you already have."

The gap is the interesting part, because it splits open weights into two things that used to be one. There is the license — the right to own, inspect, fine-tune, and redeploy the model — and there is the deployment, an infrastructure decision with its own budget. For a small or mid-sized model those arrived as a bundle. At the open frontier they no longer do. Most teams will consume K3 the way they consume a closed model: through its OpenAI-compatible API. A smaller set who need control will rent or build the GPU node to host it themselves. The license is portability insurance; running it is a separate call.

The practical move is to decide which of those two things you actually need before the parameter count scares you off. If you want frontier coding at a good price, K3's API gets you there today with no hardware at all. If you want data staying inside your perimeter, a model you can fine-tune, or insulation from one vendor's pricing and availability, then the weights are the point — and the honest next question is where the node lives: a neocloud like CoreWeave, Lambda, or Together, your own cluster, or a hybrid VPC.

Either way, benchmark it on your own workload rather than the leaderboard. A Frontend Code Arena Elo says almost nothing about your codebase, your latency budget, or how the model behaves on your hardest tickets — and at this size, the cost of finding that out in production is a rack of GPUs, not an API call.

K3 is a genuine milestone: the point where open weights caught the frontier on coding and stopped being the compromise choice. But the milestone underneath it is quieter and more durable — at this scale, "open" and "self-hosted" have come apart. Treat the license and the deployment as two decisions rather than one, and K3 is among the most useful things to happen this year for teams that want to own their AI stack instead of renting it.