Kimi K3's open weights are live on Hugging Face

Moonshot AI's Kimi K3 open weights are downloadable now. Hugging Face's listing for moonshotai/Kimi-K3 shows the repository live, with the full 2.8-trillion-parameter model available for anyone to pull and self-host — closing out a promise the company had set for July 27 that this newsroom's own automated tracking had checked repeatedly through the day without finding published, as recently as roughly an hour before this piece was filed.

What's actually in the download

The model's published configuration confirms the headline architecture: 896 routed experts with 16 selected per token plus 2 shared experts, 93 layers, a 7,168-dimension hidden size, and a context window of 1,048,576 tokens — a full million. Weights ship quantized to MXFP4 (4-bit), which is how a 2.8-trillion-parameter model compresses to a roughly 1.4-terabyte download rather than something several times larger. K3 is also natively multimodal, built around two architectural changes Moonshot calls Kimi Delta Attention and Attention Residuals — both aimed at keeping information intact as it moves through a deeper, longer-context model than K2's.

The 'active parameters' number nobody's actually confirmed

Nearly every writeup of Kimi K3 repeats a version of the same figure: roughly 50 billion active parameters per token, derived by multiplying the per-token expert count against the model's per-expert dimensions. That number is a community calculation from the published routing ratio, not a figure Moonshot itself has confirmed in an official technical report. It's a reasonable estimate — independent analyses converge on the same rough range — but readers citing '2.8T-A50B' as a vendor spec should know the 'A50B' half is outside math, not an inline Moonshot claim.

A 25-cent pelican, and what it's still good for

Independent developer Simon Willison ran his standard "draw an SVG of a pelican riding a bicycle" test against K3 the day it was first announced. His own line on the result:

That pelican took 95 input tokens and 16,658 output tokens (13,241 were reasoning tokens), for a total cost of 25 cents! — Simon Willison

He noted the model currently ships with only one reasoning-effort setting, "max" — and that the cost shows it. Willison is candid that the pelican test itself has stopped tracking real capability closely (he points to GLM-5.2 out-drawing supposedly stronger models on the same prompt), but he still runs it on every new release because it's a cheap, fast read on token cost and output validity before anyone touches the harder question of agentic tool use.

Where K3 sits against the field

AI researcher Nathan Lambert's read, published alongside the July 16 announcement, is the more consequential framing: K3 is, in his words, the closest an open-weight model has been to the frontier since DeepSeek R1, landing #3 on the Artificial Analysis Intelligence Index — behind Claude Fable 5 and GPT-5.6 Sol, ahead of xAI's Grok 4.5, and by Lambert's own count ahead of several Western labs' latest releases, including Google DeepMind's. Lambert's headline number: the gap between the best open-weight models and the best closed ones has narrowed from a widely cited 6-to-9 months to something closer to 3-to-5. Moonshot's own claimed efficiency gain — roughly a 2.5x improvement in overall scaling efficiency over K2 — is the company's figure, not independently audited, but it's consistent with Lambert's broader point: this is systematic scaling execution, not a one-off benchmark win.

The license's real fine print

The Kimi K3 License, published alongside the weights, mirrors the structure Moonshot used for K2: broad rights to use, modify, fine-tune, and redistribute the model, with two revenue-tied conditions. A company running a Model-as-a-Service business built on K3 that clears $20 million in aggregate revenue over any trailing 12 months must negotiate a separate commercial agreement with Moonshot. Separately, any product built on K3 that crosses 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3" prominently in its interface. Below those thresholds, commercial use — including hosting, fine-tuning, and redistribution — is unrestricted.

Landing in the middle of a provenance fight

The timing puts real weights into a dispute that, until now, has been argued entirely on benchmark numbers and unverified claims. The White House has accused Moonshot of distilling Anthropic's Fable models to build K3, an allegation Treasury Secretary Scott Bessent has said could trigger sanctions — one neither official has backed with public evidence, and one Moonshot has not responded to on the record. Separately, Artificial Analysis has already given K3 an independent Intelligence Index score of 57, third overall and ahead of Claude Opus 4.8. Lambert's own analysis argues the opposite of the White House's theory: he reads K3's efficiency gains as evidence of genuine architectural scaling work, not evidence that distillation explains the result — though that is one independent analyst's interpretation, not a resolution of the underlying accusation. Full published weights don't settle the dispute by themselves, but they do give outside researchers something they didn't have with the benchmark-only release: the actual architecture and parameters to inspect directly, rather than Moonshot's word for what's inside.