Moonshot's 2.8T Kimi K3.1 Pro Puts 104B Active Experts and a 1M Context Window on Hugging Face

Moonshot AI has published open weights for Kimi K3.1 Pro, a natively multimodal mixture-of-experts model the company calls its most capable 3-trillion-class release, with a one-million-token context window and post-training aimed at long-horizon coding, 3D creation, and financial research. The Hugging Face card for moonshotai/Kimi-K3.1-Pro and Moonshot’s earlier product note describe a sparse MoE with 2.8 trillion total parameters and 104 billion activated per token across 93 layers (one dense). Attention mixes 69 KDA layers with 24 Gated DeepSeek Sparse Attention layers; the MoE block routes across 896 experts with top-16 selection plus two shared experts. The vision path is MoonViT-V2 at 401M parameters. Weights ship with MXFP4 quantization-aware training and MXFP8 activations. Serving notes point to vLLM, SGLang, and TokenSpeed, plus an OpenAI/Anthropic-compatible API at platform.kimi.ai under kimi-k3.1-pro.

The repository’s initial commit is dated November 2, 2026 and contains all 96 weight shards; its November 5 update changes encoding code. Those timestamps establish repository history, not when the repository first became public. The actual public weight-release date remains unconfirmed.

The card frames K3.1 Pro as the first model in the K3.1 series, post-trained from Kimi K3 with joint reinforcement learning across hundreds of downstream domains. Reasoning effort is exposed as a top-level reasoning_effort field (low, high, max; default max). Multi-turn and tool loops require returning the full assistant message—including reasoning_content and tool_calls—so speculative thinking stays in the trajectory.

Terminal, CAD, and agent harness numbers on the card

Moonshot’s evaluation table (max effort, temperature 1.0, top-p 0.95) puts K3.1 Pro at 65.7 on Terminal-Bench 4.0 under a Claude Code harness, ahead of the GPT-6 Astra and Claude Fable 5.1 rows Moonshot lists and a point behind Claude Opus 5.5’s 66.4. Other coding cells include SWE Marathon v1.1 51.7, DeepSWE v1.1 75.5, BenchCAD with tools 92.7, and internal Kimi Code Bench 3.0 Long-horizon 61.2. Agentic rows highlight HLE-Full with tools 70.0, BrowseComp 92.1, Agents’ Last Exam (ALL) 43.2, OfficeQA Pro v2 64.4, OSWorld V2.1 79.6, and APEX-Agents 69.4. Multimodal and knowledge cells include GPQA-Diamond 94.2, MMMU-Pro 84.1, and CharXiv RQ 89.0. Cybersecurity rows are more mixed: ExploitBench 74.7 and ExploitGym 20.7 trail several closed peers Moonshot cites. Those figures are Moonshot’s harness and footnotes; several cells are internal or use company-chosen agent stacks, so treat them as vendor-reported rather than a single independent board.

License gates for large MaaS and mega-products

The Kimi K3.1-Pro License (copyright 2026 Moonshot AI) grants broad use, modification, fine-tuning, distribution, and sale, subject to notice and compliance. Two commercial gates matter for builders. Section 2 defines Model as a Service as third-party control over inference or fine-tuning inputs; if a licensee’s aggregate MaaS-related revenue exceeds US$20 million over any consecutive twelve months, a separate Moonshot agreement is required before commercial use of the Software or derivatives. Section 3 requires prominent “Kimi K3.1-Pro” UI attribution when a commercial product exceeds 100 million monthly active users or US$20 million in monthly revenue. Internal use and access through Moonshot’s official products or certified partners are carved out. That is not Apache or MIT: self-hosting and smaller commercial apps stay comparatively open; hyperscale MaaS and mega-consumer products hit contractual and branding lines.

MXFP4 serving still requires substantial infrastructure

Shipping a 2.8T multimodal agentic checkpoint with documented 1M context, MXFP4 serving paths, and a public license text gives labs a frontier-class MoE they can inspect, fine-tune, and compare against other sparse giants without waiting for a closed API. The practical package is still heavy—104B active per token is not a laptop model—and the revenue/MaaS clauses need legal review before any public inference business. For teams that can absorb the serve cost, K3.1 Pro is less a chat novelty than an open agentic coding and tool-use backbone with explicit effort controls and a published comparison table against both open and closed peers.

Moonshot’s October product note also advertised a Highspeed variant at roughly 2–3× output speed on kimi.com, the Kimi apps, Kimi Work, Kimi Code, and the API; the Hugging Face drop is the weight path for self-hosting and third-party engines. Builders who need an instruction-preserving agent loop should follow the card’s guidance on returning reasoning content; builders who only need completions can treat the MoE math and license as the release’s durable facts.