Kimi K3.1 Pro Weights Land on Hugging Face, Closing Moonshot's November 4 Promise
Moonshot AI's Kimi K3.1 Pro weights went public on Hugging Face on November 4, 2026, fulfilling the company's October 23 promise: a 2.8-trillion-parameter mixture-of-experts model with 104 billion activated parameters per token, a 1-million-token context window and a custom Kimi K3.1 Pro license, after checks through 15:18 UTC the same day still found the repository closed.
- 01THE PROMISE — Moonshot announced Kimi K3.1 Pro on October 23, 2026, shipping on its API while promising the full weights by November 4.
- 02THE CONFIRMATION — Direct Hugging Face API checks found the repository still closed at 15:18 UTC on November 4, then public with weight files and a model card by 15:55 UTC.
- 03THE WATCHPOINT — Whether independent labs reproduce Moonshot's benchmark claims on the public weights, and how the custom license is received by commercial deployers.
WHAT HAPPENEDWhat Happened
Moonshot AI's Kimi K3.1 Pro weights are public on Hugging Face. The repository moonshotai/Kimi-K3.1-Pro, which returned a 401 and was absent from Moonshot's public organization listing in checks through 15:18 UTC on November 4, 2026, was open and downloadable by 15:55 UTC the same day, closing the company's October 23 promise that the full weights would be released by November 4.
The release turns the model from an API-only product into a downloadable one. Kimi K3.1 Pro shipped on October 23 on the Kimi app, Kimi Code and the Kimi API, but its weights stayed private for twelve days while trackers and this site's own rounds kept finding the repository closed. The public repository carries the full model card, a custom Kimi K3.1 Pro license, and weight files split across 96 safetensors shards.
The model itself is Moonshot's largest open-weight release to date: a mixture-of-experts design with 2.8 trillion total parameters and 104 billion activated per token, 896 experts with 16 selected per token, 93 layers, and a 1,048,576-token context window, with native text and image input and MXFP4-quantized weights. Moonshot recommends vLLM, SGLang and TokenSpeed as inference engines for self-hosting.
WHY IT MATTERSThe deadline was the story, and Moonshot met it
Open-weight promises usually slip quietly. Moonshot put a date on this one, November 4, in its launch blog, and the repository flipped from closed to public inside that day, after morning checks still found it returning 401. A dated promise kept on the day is rare enough in this cycle to be the news itself, and it sets the next test: whether independent reproductions of the headline scores follow now that anyone can download the weights.
The timing also matters for the open-weight leaderboard. Until this release, the strongest frontier-class models of the autumn cycle were API-only, and the best downloadable vision-capable models sat well below this class. Kimi K3.1 Pro is now the largest open-weight model available by total parameters, and the first in Moonshot's K3.1 series.
WHY IT MATTERSThe scores are strong, and one of them comes with a disclosure
Moonshot's own model-card table puts Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, above GPT-6 Astra at 58.2% and just below Claude Opus 5.5 at 66.4%, with 70.0% on HLE-Full with tools and 92.1% on BrowseComp. Those are vendor-reported numbers, measured at maximum reasoning effort, and they should be read as Moonshot's claims until outside labs rerun them on the public weights.
The more unusual disclosure is in cybersecurity. Moonshot says it removed cybersecurity-related data during mid-training and reinforcement learning for the open-weight version, and reports that an unsanitized checkpoint scored 90.5% on its internal xdev-bench-hard test against 81.0% for the released version. The open weights are therefore deliberately less capable at offensive security work than the model Moonshot itself evaluated, a trade the company made under China's export restrictions. Builders measuring the public model against the API model should expect that gap.
SIGNAL JUDGMENT — ANALYSISA dated open-weight promise, kept on the day
Tokenta Signal's judgment: the Kimi K3.1 Pro weights release matters less as another benchmark table and more as a deadline met in public. Moonshot named November 4, and the repository flipped from closed to downloadable inside that day. The claim to discount is the vendor scoreboard, which now faces its first real test, since anyone can rerun it. The claim to watch is the sanitized cybersecurity capability, where Moonshot openly ships a weaker open model than the one it evaluated, and the license terms, which are custom rather than permissive.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Moonshot AI's Hugging Face repository for Kimi K3.1 Pro was opened and verified directly: a public model page describing an open-weight, natively multimodal agentic model with 2.8 trillion total parameters, 104 billion activated per token and a 1-million-token context window, with weight files, a custom Kimi K3.1 Pro license and deployment guidance for vLLM, SGLang and TokenSpeed. Moonshot's October 23 tech blog supplies the November 4 weights promise that this release fulfills.
PRIMARY SOURCESPrimary Sources
- Hugging Face: moonshotai/Kimi-K3.1-Pro
https://huggingface.co/moonshotai/Kimi-K3.1-Pro - Moonshot AI: Kimi K3.1 Pro tech blog
https://www.kimi.ai/blog/kimi-k3-1-pro
RELATEDRelated Intelligence
- Canva Restructures 22% of Its Workforce as AI Serving Costs Bite: 1,271 Affected, Growth Forecast Cut to 20%
- GlobalFoundries Posts Record 31.4% Margin as AI Packaging Deals Stack Up: $2B TSMC Pact, $21B U.S.-Japan Fabs
- Lambda Pledges $33.1 Billion for Alabama AI Campus, Revealing Itself as Project Red Clay's Tenant
- Google Pledges $15 Billion for Malaysia Data Centers, With Clean Energy and Skills Attached