Moonshot AI · Release tracker

Kimi K3 open weights

Moonshot AI ran two launches for one model. Kimi K3 went live on July 16, 2026 through the Kimi app, Kimi Work, Kimi Code and the API — but the weights, the part that makes “open-weight” mean anything, are promised separately for July 27, 2026 — now two days away. They have not shipped and the license still has not been named.

Checked 2026-07-25: Moonshot's Hugging Face organisation lists 18 models, the newest of them Kimi-K2.7-Code. There is no K3 repository, gated or otherwise. This page is re-checked on every radar pass and will be updated the day the weights land — or the day the date passes without them.

Current status

Model live
Yes — since July 16, 2026
Weights downloadable
No
Promised date
July 27, 2026

Artificial Analysis still classifies K3 as proprietary, with weights “not publicly available” — on the straightforward grounds that there is nothing public to download. Moonshot describes it as an open 3-trillion-parameter-class model. Both are accurate descriptions of different things — an intention and a state — and the distinction matters if you are planning around it.

Where we looked, on 2026-07-25: huggingface.co/moonshotai — the only place an official release can appear, since every prior Kimi weight drop landed there. Newest entry: Kimi-K2.7-Code. Note that third-party repositories using the K3 name already exist on Hugging Face; none of them are Moonshot's, and none of them carry Moonshot's weights.

What is confirmed

Confirmed
Released
July 16, 2026, via API and Moonshot's own apps
Total parameters
2.8 trillion
Context window
1M tokens
API price
$3.00 in / $15.00 out per 1M
Modality
Native multimodal input (text and images); text-only output
Reasoning
Always-on thinking mode — it cannot be disabled
Intelligence Index
57 — #7 of 190 on Artificial Analysis as of July 25 (it entered at #4)
Frontend Code Arena
#1, around 1,679 Elo — ahead of Claude Fable 5
Where to use it today
Moonshot's first-party API, and OpenRouter

One of those numbers has already moved, which is worth saying out loud. Launch coverage from July 16–21 put K3 fourth on the Artificial Analysis Intelligence Index, and that is still what most of it says. Its score has not changed — 57 — but four models have entered the index since, and K3 now sits seventh of 190. Nothing about the model got worse. The field moved underneath a fixed number, and a benchmark rank is a position, not a property.

What is only promised

Unconfirmed

  • The July 27, 2026 weight release itself.

    It is a stated plan from Moonshot, not a shipped artifact. There is no repository to watch, no gated form to join, and no staged partial release.

  • The license.

    “Modified MIT” is the widely circulated expectation, and it is a reasonable one: we checked, and both Kimi-K2.6 and Kimi-K2.7-Code ship on Hugging Face under a Modified MIT License covering code and weights alike. But precedent is not publication. Moonshot has released no license text for K3, and a modified MIT is not MIT — the modifications are the whole story, and for K3 there is nothing to read yet.

  • Architecture details.

    Kimi Delta Attention — credited with up to 6.3× faster decoding — plus Attention Residuals and a Stable LatentMoE that routes 896 experts, activating 16 of them at roughly 50B parameters per token. All of this comes from community analysis. Expect the authoritative version in the model card, not before.

  • Quantization format.

    MXFP4 weights with MXFP8 activations, quantization-aware from the fine-tuning stage, is the reported plan. If it holds it is the difference between 1.4 TB and something far worse.

What “you can self-host it” actually costs

This is the part that gets skipped. Open weights on a 2.8 trillion parameter model do not mean you can run it — they mean a well-capitalised infrastructure team can run it. At MXFP4 the weights alone occupy roughly 1.4 TB. A practical serving cluster starts around eight nodes of 8×80GB GPUs, about 5.12 TB of aggregate memory once you account for activations, KV cache across a 1M-token context, and routing overhead on 896 experts.

Treat any smaller download figure with suspicion until the model card lands. Numbers in the 500–600 GB range are circulating for “K3 in BF16”, and they do not survive the arithmetic: at two bytes per parameter, 2.8 trillion parameters is about 5.6 TB, not 600 GB. Figures that size belong to the 1-trillion-parameter K2 generation. The only number that gets K3 near 1.4 TB is the MXFP4 quantization, and that is still a reported plan rather than a shipped file.

You also need an inference stack that understands mixture-of-experts scheduling — vLLM, TensorRT-LLM or SGLang — rather than whatever you use for a 70B dense model. None of this fits on a workstation, a single H100 box, or a consumer GPU, and any page telling you to “run Kimi K3 locally” on desktop hardware is describing something that will not happen at full precision.

So who is the weight release actually for?

Three groups. Inference providers, who will host it and compete on price against Moonshot's own $3 / $15 rate. Enterprises with a hard data-residency requirement, for whom the cluster cost is cheaper than the compliance problem. And researchers, who will distil and quantise it down into things that do fit on smaller hardware — which is historically where most of the downstream value of an open-weight release ends up. For everyone else, the practical consequence of July 27 is that the API price falls.

Why the date is under pressure

Three days after K3 shipped, Alibaba previewed Qwen3.8-Max — 2.4 trillion parameters, also promising open weights, also without a date. Reporting has read the timing as an attempt to blunt K3's momentum, and it puts Moonshot in an awkward spot: it is currently the lab charging API-only rates for a model it calls open, while its closest competitor markets openness as the differentiator.

That is a reason to expect Moonshot to hit the date rather than slip it. It is not evidence that they will. Weight releases slip routinely, usually for safety review or license drafting rather than engineering, and neither of those produces a public warning first.

What we will change here, and when

  • If the weights ship on July 27, 2026: this page gets the repository link, the actual license text, the model card's activated-parameter count, the real file sizes, and the first third-party hosted prices as they appear.
  • If the date passes quietly: we say so, with the date we checked. A promised release that has slipped is a fact worth publishing, and it is the one thing the pages racing to rank on this keyword will not go back and do.
  • Either way: the pricing side moves to Kimi K3 pricing, which tracks the API rate and the cost-per-task numbers.

Sources

Figures on this page last checked against these sources on 2026-07-25. Vendors change pricing and specs without notice — if a number here disagrees with the vendor's own page, trust the vendor.