Moonshot AI released the model weights of Kimi K3.1 Pro on Hugging Face on 4 November, posting on X at 16:00 UTC: “Releasing the model weights of Kimi K3.1 Pro”. The model card lists 2.8T total parameters, 104B activated, and a context length of 1,048,576 tokens.
In the benchmark table on that card, Kimi K3.1 Pro scores above Claude Opus 5.5 on 16 of the 37 non-internal rows where both have a score, by our count. Its largest lead is 5.4 points, and Opus 5.5 is ahead by more than 20 points on three rows, by our arithmetic.
Business Pill · Open Weights Explained
A short Business Engineer explainer of open weights: rent the model or run it yourself. It teaches the general idea only and does not discuss Moonshot or Kimi K3.1 Pro.
The key insight: In Moonshot’s own table, Kimi K3.1 Pro is ahead of Claude Opus 5.5 on 16 of 37 non-internal rows, by our count, with leads of at most 5.4 points and gaps of up to 33.5 points; the weights are open, but Model as a Service hosts with more than $20 million in 12-month revenue need a separate agreement with Moonshot.
Moonshot calls K3.1 Pro “an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date”. The card lists a Mixture-of-Experts design with 896 experts, 16 selected per token, and “MXFP4 weights / MXFP8 activations (quantization-aware training)”. By our arithmetic, 104B of 2.8T parameters, or 3.7%, are activated for each token.
The repository holds 96 weight files totalling about 1.56 TB, by our count of the Hugging Face file listing. The card names vLLM, SGLang and TokenSpeed as recommended inference engines, and the model stays available through Moonshot’s own API as kimi-k3.1-pro.
Kimi K3.1 Pro score minus Claude Opus 5.5 score on each of the 35 point-scored, non-internal rows of Moonshot’s benchmark table. Source: Moonshot AI, Kimi K3.1 Pro model card on Hugging Face, read 4 Nov 2026. Scores are as reported by Moonshot; the differences and counts are our arithmetic.
Moonshot AI posted the release at 16:00 UTC on 4 November, linking to the Hugging Face repository. The model card there lists 2.8T total parameters and 104B activated; by our count, the repository’s 96 weight files total about 1.56 TB.
Moonshot AI just released the Kimi K3.1 Pro weights on Hugging Face.
The model has 2.8 trillion parameters, and the weights alone reportedly take up around 1.56 TB.
Want to run it locally One possible setup is a server with 8 NVIDIA B300 GPUs and 2.3 TB of GPU memory. Estimated hardware cost: $550,000–$650,000.
@kimmonismus posted at 16:27Z that the weights reportedly take up around 1.56 TB, in line with our count of the repository, and that one possible setup for running the model locally is a server with 8 Nvidia B300 GPUs at an estimated $550,000 to $650,000. The setup and cost are the post’s estimates; we have not checked them. The 2.8 trillion parameter figure is also on Moonshot’s model card.
What the Licence Allows
The weights ship under the Kimi K3.1-Pro License, not a standard open-source licence. It grants the rights “to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software”, subject to conditions.
The main condition is for hosts. If a licensee operates a “Model as a Service” business and its revenue with affiliates “exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months”, the licensee “must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose”.
A second condition covers large products: above “100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue”, the name “must be prominently displayed on the user interface”. Neither condition applies to internal use, which the licence defines as use that does not make the software, its outputs or its capabilities available to third parties.
What Moonshot’s Table Shows
The card compares Kimi K3.1 Pro with Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3 across 47 benchmark rows, eight of them labelled internal. Excluding those eight, Kimi K3.1 Pro and Opus 5.5 both have scores on 37 rows, and Kimi K3.1 Pro is ahead on 16, by our count.
The leads are small: the largest is 5.4 points, on SWE Marathon v1.1, 51.7 against 46.3, by our arithmetic. The gaps the other way are larger: 33.5 points on Program Bench, 25.2 on Terminal-Bench-Science 0.1 and 21.5 on HLE-Diamond, by our arithmetic. On the two cyber security rows, ExploitBench and ExploitGym, Kimi K3.1 Pro trails Opus 5.5 by 16.3 and 13.8 points, by our arithmetic.
Moonshot’s tech blog said in October: “For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness”. The card says Kimi K3.1 Pro was run at reasoning effort max; the Opus 5.5 column is labelled “max, w/ fallback”, and the card’s footnotes say many comparison scores come from other labs’ reports or public leaderboards.
Two Chinese Labs, One Week
Also on 4 November, Tencent released Hy4 through its API, and its release page says: “We plan to open-source Hy4’s model weights”. It gives no date.
OpenRouter’s public listing, read at 16:20 UTC, shows Kimi K3.1 Pro at $2 per million input tokens and $10 per million output tokens, half of the $4 and $20 it lists for Claude Opus 5.5. Tencent’s list price for Hy4 output is $2.501, so Kimi K3.1 Pro’s output price is 4.0 times Hy4’s, by our arithmetic.
The two labs’ tables differ on similarly named tests. On APEX-Agents, Moonshot’s card gives Kimi K3.1 Pro 69.4 and Opus 5.5 73.5; on APEX-Agents v1.1, Tencent’s table gave Kimi K3.1 Pro 68.2 in Tencent’s testing and listed Opus 5.5 at 73.5 and 74.8.
On demand, a Menlo Ventures report, which we covered on 3 November, put Kimi at 61% of open-source model spend among the U.S. enterprises it surveyed. Menlo is an investor in Anthropic and OpenRouter.
As we read the licence, Moonshot opens the weights to most users while keeping a contract with the largest model-API hosts: the separate-agreement condition applies above $20 million in revenue over any 12 months.
As we read the table, 12 of the 16 rows where Kimi K3.1 Pro is ahead are in its Agentic section, by our count, while the largest gaps are in Program Bench, Terminal-Bench-Science, HLE-Diamond and the two cyber security rows, where Moonshot said it removed cyber-related training data for the open-weight release.
Moonshot AI, Kimi K3.1 Pro model card, Hugging Face
“Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date”
Three Implications
FOR MODEL HOSTS The licence requires a separate agreement with Moonshot for Model as a Service businesses above $20 million in revenue over any 12 months.
FOR API BUYERS OpenRouter lists Kimi K3.1 Pro at $10 per million output tokens, half of Claude Opus 5.5’s $20 and 4.0 times Tencent’s $2.501 list price for Hy4, by our arithmetic.
FOR BENCHMARK READERS On APEX-Agents, Moonshot’s card gives Kimi K3.1 Pro 69.4, while Tencent’s table gave it 68.2 in Tencent’s testing.
The framework says: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”
As we read it, Moonshot’s licence draws a line of that kind: the weights are open to most users, while the largest model-API hosts need a separate agreement. The framework predates this release and does not discuss Moonshot or Kimi K3.1 Pro.
The benchmark scores are Moonshot’s own, run or collected by Moonshot; we have not seen independent results on the released weights. The card does not say how much hardware is needed to serve the model, and the licence does not say what terms the separate agreement for large hosts carries. Tencent has given no date for the Hy4 weights.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Moonshot’s Kimi K3.1 Pro weights, 2.8T parameters in about 1.56 TB of files by our count, are on Hugging Face under a licence that requires model-API hosts with more than $20 million in 12-month revenue to sign a separate agreement. In Moonshot’s own table the model is ahead of Claude Opus 5.5 on 16 of 37 rows, by our count, at half Opus 5.5’s list price.
A note on sourcing. Model details, benchmark scores and licence terms come from Moonshot AI’s Kimi K3.1 Pro model card and LICENSE file on Hugging Face, read on 4 November 2026; scores are Moonshot’s own. Prices are OpenRouter’s public listing at 16:20 UTC and Tencent’s release page as reported in our earlier piece. Counts, differences and ratios are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
