Moonshot ships Kimi K3.1 Pro: 2.8T MoE with 1M context

Moonshot AI's new flagship, Kimi K3.1 Pro, is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion activated per token, and it ships with a 1,048,576-token context window. The model card on Hugging Face describes it as "an open-weight, natively multimodal agentic model and the most capable 3T-class model from Moonshot AI to date."

Under the hood sit 93 layers, 896 experts with 16 selected per token plus 2 shared, and an attention stack listed as "69 KDA + 24 Gated DSA". The card says "the model replaces full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling compared to Kimi K3." Weights are quantization-aware trained to MXFP4 with MXFP8 activations, which keeps a 3T-class model closer to a practical inference footprint. A 401M-parameter MoonViT-V2 handles the vision side.

Moonshot says K3.1 Pro was post-trained from Kimi K3 "with large-scale joint reinforcement learning across hundreds of downstream domains," pitching it for coding and agent workloads. Self-reported scores include Terminal-Bench 4.0 at 65.7, SWE Marathon v1.1 at 51.7, BrowseComp at 92.1, HLE-Full with tools at 70.0, and GPQA-Diamond at 94.2. Several of the listed rows are internal benchmarks (Kimi Code Bench, DECK-Arena, Business Operation), so the externally comparable scores are the ones to weight first; there is no third-party evaluation attached.

The weights sit under a bespoke "Kimi K3.1 Pro License" rather than Apache 2.0 or MIT, and API access runs through platform.kimi.ai with OpenAI- and Anthropic-compatible endpoints. Two of the researchers we track posted the Hugging Face link within hours.