Kimi K3.1 Pro

How it is built

Kimi K3.1 Pro is an open-weight mixture of experts from Moonshot AI, announced on 24 October 2026, with 2.8T parameters and 104B of them active for each token.¹¹1Moonshot AI, Kimi K3.1 Pro model card, November 2026. Evaluation settings: reasoning effort max, temperature 1.0, top-p 0.95. https://huggingface.co/moonshotai/Kimi-K3.1-Pro ²²2Moonshot AI, Kimi K3.1 Pro tech blog, listed on the Kimi blog index as 24 October 2026. https://www.kimi.ai/blog/kimi-k3-1-pro It accepts text and image input and has a context of 1,048,576 tokens.¹¹³³3Moonshot AI, Kimi K3.1 Pro config.json: first_k_dense_replace 1; DSA layers 4, 8 and every fourth layer to 92, plus 93; max_position_embeddings 1048576; 896 experts, 16 per token, 2 shared. https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/config.json

The model has 93 layers, and only the first is dense. The other 92 route each token to 16 of 896 experts, with 2 shared experts always on.¹¹³³

Of the 93 layers, 69 use Kimi Delta Attention (KDA), a linear form of attention. The other 24 use Gated DeepSeek Sparse Attention (Gated DSA) and sit at every fourth layer, with one more at the top.¹¹³³

Attention runs at a hidden size of 7,168 with 96 heads. The vocabulary holds 160K tokens.¹¹ Images pass through MoonViT-V2, a vision encoder of 401M parameters.¹¹

What is new

Moonshot describes K3.1 Pro as a post-trained version of Kimi K3, with much more post-training compute. It is the first model of the K3.1 series.²²

The architecture matches K3 in size and expert layout. The one change Moonshot lists is that the Gated MLA layers of K3 became Gated DSA layers, for long-context modelling.¹¹

On Terminal-Bench 4.0 Moonshot reports 65.7 for K3.1 Pro, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra.²² On its internal Kimi Code Bench 3.0 Long-horizon it reports 61.2, against 55.95 for K3.²² Settings are reasoning effort max, temperature 1.0 and top-p 0.95.¹¹

Moonshot released a faster variant at the same time, Kimi K3.1 Pro Highspeed, at 2 to 3 times the speed.²²

Running it

The weights were published on Hugging Face on 2 November 2026.⁴⁴4Hugging Face, moonshotai/Kimi-K3.1-Pro commit history: "Initial commit" on 2 November 2026 at 04:13 UTC. https://huggingface.co/moonshotai/Kimi-K3.1-Pro/commits/main The checkpoint uses MXFP4 weights with MXFP8 activations, from quantization-aware training.¹¹

Under the Kimi K3.1-Pro License, a company that runs a Model as a Service business needs a separate agreement for commercial use once its revenue passes US$20 million over twelve months.⁵⁵5Moonshot AI, Kimi K3.1-Pro License, sections 2 and 4. https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE Internal use is exempt.⁵⁵

On the Kimi API, kimi-k3.1-pro costs $2.00 per million input tokens and $10.00 per million output tokens. A cache hit costs $0.20.⁶⁶6Kimi API Platform, chat model pricing, per million tokens, read 6 November 2026. https://platform.kimi.ai/docs/pricing/chat The Highspeed variant costs $4.00 for input and $20.00 for output.⁶⁶