Moonshot open-sources Kimi K3.1 Pro, a 2.8T-parameter multimodal agentic model with 1M-token context

Moonshot AI has published the weights for Kimi K3.1 Pro on Hugging Face, turning a frontier-scale model into a freely downloadable artifact. The release, which hit the Hacker News front page on 4 November, is the first model in the K3.1 series. According to the model card, it is the strongest model Moonshot has produced in the roughly three-trillion-parameter class, and it is built as an agentic system that works across text and images.

What Moonshot shipped

The model card says K3.1 Pro was post-trained from Kimi K3 using joint reinforcement learning at large scale across hundreds of downstream domains, with claimed gains in coding, 3D creation, financial research and factual accuracy. The headline specifications:

• Mixture-of-Experts architecture with 2.8 trillion total parameters and 104 billion activated per token
• 93 layers, with 896 experts of which 16 are routed per token alongside 2 shared experts
• A context window of 1,048,576 tokens
• A 401-million-parameter vision encoder, MoonViT-V2, accepting text and image input
• Weights distributed in MXFP4 with MXFP8 activations, the result of quantization-aware training

The attention design marks a departure from Kimi K3: the card explains that full-attention gated MLA layers have been replaced with 24 Gated DeepSeek Sparse Attention layers, paired with 69 KDA layers, a change Moonshot credits with more efficient long-context modeling. The sparse-attention index uses 32 heads and selects a top-k of 2,048 tokens. The vocabulary covers 160,000 tokens and the activation function is SiTU-GLU.

How it scores

Moonshot's evaluation table pits K3.1 Pro against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3, with proprietary rivals run at maximum effort and permitted fallback behavior.

The results are mixed but competitive at the top end. On SWE Marathon v1.1 the model posts 51.7, ahead of Fable 5.1 (45.6) and the 42.5 scores of both GPT-6 Astra and GLM-5.3. It reaches 65.7 on Terminal-Bench 4.0, within a point of Opus 5.5 and ahead of the other competitors listed. On the browsing benchmark BrowseComp it scores 92.1, slightly above GPT-6 Astra (91.5) and Opus 5.5 (90.8). Its 70.0 on HLE-Full with tools is the best in that row, and GPQA-Diamond lands at 94.2 against 96.1 for GPT-6 Astra.

The gaps run the other way in several places. Terminal-Bench-Science shows 38.1 versus 68.1 for GPT-6 Astra, Program Bench shows 31.5 versus 65.0 for Opus 5.5, and on HLE-Diamond the model's 39.4 trails GPT-6 Astra's 65.8. Cybersecurity benchmarks are a clear weak spot: ExploitBench scores 74.7 against a reported 100 for GPT-6 Astra, and ExploitGym yields 20.7 versus 42.4. On an internal hallucination benchmark run with tools, where lower is better, K3.1 Pro records 8.3 — better than most rivals but behind GPT-6 Astra's 5.5.

Reading the numbers carefully

These are vendor-run evaluations, and the footnotes deserve attention. Several benchmarks are labeled internal, including the Kimi Code Bench suites, DECK-Arena, Design Bench, Business Operation and xdev-bench-hard. Kimi's own coding runs were evaluated with the Claude Code harness, while competitor figures come from a mix of official leaderboards, system cards and vendor release blogs. A dash in the table simply means no score was reported, so column-level comparisons are not always complete.

Why it matters

A 2.8-trillion-parameter model with a million-token context and native image understanding, released as open weights, narrows the distance between downloadable models and API-only frontier systems. Even with only 104 billion parameters active per token and 4-bit weights, running it locally is a serious hardware undertaking, but research groups can now inspect, fine-tune and distill a system of this class rather than treat it as a black box. The release also continues the pattern of Chinese labs — Moonshot alongside DeepSeek and Zhipu's GLM line — shipping open weights that compete directly with proprietary offerings, and the adoption of sparse attention signals where efficient long-context design is heading. Until independent evaluations appear, Moonshot's benchmark claims should be treated as promising rather than settled.