Kimi K3.1 Pro Open Weights: 2.8T Parameters, 65.7% on Terminal-Bench 4.0

Moonshot AI has published the weights of Kimi K3.1 Pro on Hugging Face, making a 2.8-trillion-parameter model downloadable by anyone with the storage to hold it. Moonshot's launch post promised the full weights "by Nov 4, 2026", and the repository on Hugging Face is dated 2 November. On Moonshot's own numbers, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5, 58.2% for GPT-6 Astra and 57.9% for Claude Fable 5.1. If those numbers hold up under independent testing, the gap between this open-weight model and the best closed model in Moonshot's Terminal-Bench table is under a point. The release also comes with an unusual admission: Moonshot says it deliberately trained the model to be worse at offensive cyber work than it could have been.

What happened

K3.1 Pro is a post-trained upgrade of Kimi K3, and the first model in the K3.1 series. The model card describes a Mixture-of-Experts with 2.8 trillion total parameters and 104 billion activated per token: 896 experts, 16 selected per token, plus 2 shared experts. The context window is 1,048,576 tokens. Moonshot replaced K3's full-attention layers with what it calls Gated DeepSeek Sparse Attention for cheaper long-context work, and trained with quantisation awareness so the released weights are MXFP4 with MXFP8 activations.

Moonshot's launch post says K3.1 Pro and a K3.1 Pro Highspeed variant, which it claims runs 2 to 3 times faster, are available in Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The API model ID is kimi-k3.1-pro, with OpenAI- and Anthropic-compatible endpoints. Moonshot's card recommends vLLM, SGLang or TokenSpeed for self-hosting.

The cyber part is in the launch post. Moonshot says it identified and removed cyber-related data from both training stages, and that "K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce". On its internal xdev-bench-hard, the post says data removal took the score from 90.5% to the released model's 81.0%. Moonshot also says it intends to raise the cyber capability in open weights progressively, and says K3.1 Pro's cyber capability remains behind frontier proprietary models that have been publicly available for at least three months.

The numbers

All scores are from Moonshot's model card, at max reasoning effort. Moonshot ran Kimi on its own harnesses or Claude Code; comparison scores come from vendor reports and leaderboards, which use different harnesses.

Other figures worth knowing: the Hugging Face API reports 2,780,110,363,392 parameters in the safetensors files and about 1.56 TB of repository storage (queried on 10 November 2026). Moonshot's pricing page says batch inference for kimi-k3.1-pro is billed at 60% of the standard price.

Why it matters

On Terminal-Bench 4.0, the previous best open-weight scores in this table are DeepSeek V4.1 Flash at 42.3 and GLM-5.3 at 41.8. Mistral Large 4, released on 29 October, reports 41% in its launch post. K3.1 Pro's 65.7 is a jump of more than 20 points for open weights, landing next to Opus 5.5 rather than a generation behind.

The cyber decision matters as policy. Closed labs gate their most capable cyber models behind vetting programs; Anthropic's Mythos and Google's Gemini 4 Carbon both work that way. An open-weight lab cannot gate anything once the files are public, so Moonshot chose to remove capability before release instead. Whether that approach is the right standard is debatable, but it is a concrete, written policy, which is more than most open releases offer.

What sceptics say

The table that flatters K3.1 Pro on Terminal-Bench also shows where it trails. On HLE-Full without tools it scores 51.1 against 59.8 for Opus 5.5. On Program Bench it scores 31.5 against 65.0. Several of its best results are on Moonshot's internal benchmarks, and Kimi was evaluated in its own harness or Claude Code while some rivals' numbers come from their vendors' reports. Cross-harness comparisons can move by several points.

Then there is provenance. Mistral's Pierre Stock, launching Large 4 a few days earlier, told Fortune that companies should question whether Chinese open models can be trusted, citing possible export controls. That is a competitor talking, but procurement teams in regulated industries ask the same question, and "the weights are already on our disks" only answers part of it.

Finally, size. A 1.56 TB repository with 104 billion active parameters is not a model most teams will self-host casually, even with MXFP4 weights.

What it means for engineering and DevOps teams

My take: K3.1 Pro is the first open model I would put in a head-to-head eval against Opus 5.5 for agentic coding, and the eval is the point. Run it before you believe either vendor.

• Evaluate on the API first. The Kimi API is OpenAI-compatible, so an existing eval harness or gateway can target it with a base URL change.
• Preserve the reasoning content. The card says multi-turn and tool-call conversations must pass back the complete assistant message, including reasoning_content and tool_calls. Gateways that strip unknown fields will break it quietly.
• Read the license thresholds. The K3.1 Pro license is MIT-style, but a Model-as-a-Service business above $20 million in trailing 12-month revenue needs a separate agreement, and products above 100 million monthly users or $20 million monthly revenue must display "Kimi K3.1-Pro". Internal use is exempt from both.
• Migrate off K2 if you use it. Moonshot's pricing page says K2 models are closed to new users from 31 October 2026 and will be retired by the end of December 2026.
• Size the hardware honestly. If self-hosting is the goal, start from the inference engine comparison and the cost model in Self-Hosting LLMs vs API Costs.

# Check what you would be downloading before you pull 1.5 TB of weights.
curl -s "https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro?expand[]=safetensors&expand[]=usedStorage" \
  | jq '{id, params: .safetensors.total, storage_tb: (.usedStorage / 1e12)}'

FAQ

How big is Kimi K3.1 Pro?

2.8 trillion total parameters with 104 billion active per token, according to Moonshot's model card, and about 1.56 TB of storage on Hugging Face. The weights are MXFP4 with MXFP8 activations.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

On Moonshot's own numbers it is close on Terminal-Bench 4.0 (65.7 vs 66.4) and ahead on BrowseComp, but behind on HLE-Full and Program Bench. No independent evaluation has settled it yet.

Can I use Kimi K3.1 Pro commercially?

Yes, under a modified MIT license. A Model-as-a-Service business with more than $20 million in trailing 12-month revenue needs a separate agreement with Moonshot, and very large products must display the model name.

Related: Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing, LLM Inference Engines 2026: vLLM vs SGLang vs TensorRT-LLM vs Dynamo and Coding Agents in 2026: Cursor at $2B Then Sold to SpaceX, Cognition at $1B, 58% on Terminal-Bench, and the Only RCT Still Says Slower.