Moonshot AI releases full 1.56TB weights for Kimi K3.1 Pro

Summary

• Moonshot AI published Kimi K3.1 Pro's complete open weights on Hugging Face, 112 files totaling roughly 1.56 terabytes.
• The 2.8-trillion-parameter mixture-of-experts model scores 65.7 on Terminal-Bench 4.0, ahead of GPT-6 Astra's 58.2.
• Companies can now self-host and fine-tune a frontier-class coding and agent model instead of renting closed APIs.

The latest

The complete weights of Kimi K3.1 Pro are now downloadable on Hugging Face, where Moonshot AI's repository carries 112 files totaling about 1.56 terabytes, according to the platform's API. The company describes it as its strongest open-weight model in the 3-trillion-parameter class, built on a mixture-of-experts architecture with a one-million-token context window and native text and image input.

Details

• The architecture: Per the model card, K3.1 Pro runs 2.8 trillion total parameters with 104 billion active per token across 93 layers. It uses 896 experts, 16 selected per token plus two shared, a 160K vocabulary, and a context length of 1,048,576 tokens.
• The attention change: Moonshot says it replaced Kimi K3's Gated MLA full-attention layers with Gated DeepSeek Sparse Attention for more efficient long-context modeling. The configuration combines 69 KDA layers with 24 Gated DSA layers, alongside a 401M-parameter MoonViT-V2 vision encoder.
• Coding benchmarks: On Terminal-Bench 4.0 the card reports 65.7 for K3.1 Pro against 58.2 for GPT-6 Astra, 57.9 for Claude Fable 5.1 and 66.4 for Claude Opus 5.5. On SWE Marathon v1.1 it posts 51.7, the highest figure in that row.
• Where it trails: The same card shows K3.1 Pro behind rivals elsewhere: 31.5 on ProgramBench against 65.0 for Claude Opus 5.5, 39.4 on HLE-Diamond against 65.8 for GPT-6 Astra, and 74.7 on ExploitBench where one competitor scores 100.
• Agent and office tasks: Moonshot reports 92.1 on BrowseComp, 79.6 on OSWorld V2.1 and 70.0 on HLE-Full with tools, and claims a state-of-the-art 64.4 on OfficeQA Pro v2 against 54.5 and 54.4 for the two closest listed models.
• Testing conditions: All K3.1 Pro results were produced at max reasoning effort, temperature 1.0 and top-p 0.95, with Terminal-Bench evaluated through the Claude Code harness. The reasoning_effort field accepts low, high and max, defaulting to max.
• The license terms: The Kimi K3.1-Pro License permits use, modification, distribution, fine-tuning and sale. But any operator running model-as-a-service with more than $20 million in gross revenue over any 12 consecutive months must sign a separate agreement with Moonshot before commercial use.
• The attribution clause: Products exceeding 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently in the interface. Neither condition applies to internal use or to Moonshot's own products and approved inference partners.
• Running it: API access is available at platform.kimi.ai as kimi-k3.1-pro with OpenAI- and Anthropic-compatible interfaces. Moonshot recommends vLLM, SGLang and TokenSpeed, and requires full assistant messages including reasoning_content and tool_calls be resent in multi-turn exchanges.
• The 3D claim: In its October 24 announcement blog, Moonshot said the model works with Unity, Unreal Engine and Godot, and writes and runs scripts for Blender and CAD software without MCP connectors. It also launched a Highspeed variant it describes as 2–3 times faster.

Background

Moonshot announced K3.1 Pro on October 24 as the first model in the K3.1 series, post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of domains. At launch it promised full weights by November 4; the Hugging Face repository is dated November 2.

Between the lines

The revenue thresholds in the license are the commercial boundary. A company self-hosting the model internally faces no conditions at all, which is what makes the release meaningful for regional firms wanting a coding and agent model on their own servers. The restrictions bite only on resale — anyone building a competing inference business at scale must return to Moonshot and negotiate.

What's next

Watch whether independent evaluations reproduce the Terminal-Bench and OfficeQA Pro figures, and whether inference providers add the 1.56TB model to their hosted catalogs given the storage and serving costs involved.

Source: Moonshot AI (Hugging Face model card, license file), kimi.ai