Moonshot AI releases Kimi K3.1 Pro weights: 2.8 trillion parameters, 1 million token context

Weights arrive on the promised date

Chinese AI company Moonshot AI released the model weights of Kimi K3.1 Pro on Hugging Face on Nov. 4 (16:00 UTC). It unveiled the model on Oct. 23 and promised open weights “by Nov 4.” Until then, the model was available only through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The Hugging Face repository was created on Nov. 2, and the weights are split across 96 files.

According to the model card, K3.1 Pro was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of domains. Moonshot calls it “our most capable 3T-class model to date.”

Kimi.ai@Kimi_Moonshot

Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Kim…
We'll keep advancing open intelligence.

Architecture

From the model card

• Mixture of experts (MoE): 2.8 trillion total parameters, 104 billion active per token
• 93 layers and 896 experts; each token uses 16 experts plus 2 shared experts
• A context window of 1,048,576 tokens
• Text and image input, with the 401-million-parameter MoonViT-V2 vision encoder
• MXFP4 weights and MXFP8 activations, with quantization-aware training

Moonshot replaced K3’s full-attention layers with Gated DSA (DeepSeek Sparse Attention) layers for more efficient long-context modeling. The model runs on vLLM, SGLang and TokenSpeed. AMD, Nebius and Verda supported it from day one; AMD says it runs on eight Instinct MI355X GPUs.

Company benchmarks and cybersecurity

In Moonshot’s table, the model scores 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Some gaps are wide: on HLE-Diamond, K3.1 Pro scores 39.4% to Opus 5.5’s 60.9%. On an internal test, its claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot says the test is deliberately hard and does not reflect error rates in everyday use.

According to the tech blog, the company removed cyber-related data from both mid-training and RL environments for the open-weight release. On its internal xdev-bench-hard, a checkpoint trained without those measures solved 90.5% of tasks, while the released model reached 81%. Moonshot says it aims not to release open weights that extend the absolute frontier of cyber capability.

An MIT-style license with two conditions

The weights ship under Moonshot’s own “Kimi K3.1-Pro License.” Like the MIT license, it allows use, modification, distribution and sale, but it adds two conditions. Companies that offer the model as a service, for example through an API, and earn more than $20 million over 12 consecutive months must sign a separate agreement with Moonshot before commercial use. Products with more than 100 million monthly active users or more than $20 million in monthly revenue must prominently display “Kimi K3.1-Pro.” Internal use and access through Moonshot’s certified inference partners are exempt.

In its Oct. 27 Intelligence Index v5, Artificial Analysis had ranked the model the top open-weights model at 46 points, noting that its weights were not yet released. It trailed the leader, Claude Opus 5.5, by 12 points.