Moonshot AI released Kimi K3.1 Pro on October 23, 2026, an upgrade to its open-weight K3 model. Moonshot says it comes within a point of Claude Opus 5.5 on a coding-agent test. It also says it stripped hacking-related material out of training before the weights go public, and it published how much skill that cut removed.
Moonshot's announcement calls it "our most capable 3T-class open-weight model to date." The model is live now in Kimi's apps and API. Moonshot promises the full weights by November 4.
What Moonshot shipped
K3.1 Pro is not a new design. Moonshot says it took K3, released in July, and trained it further with much more compute in the post-training stage. Post-training is the phase after the base model learns language, when it is taught to follow instructions and use tools.
Moonshot calls the model "3T-class," meaning about 3 trillion parameters, the internal values a model learns. Reports differ on the exact count. SiliconANGLE describes 3 trillion parameters spread across 896 expert networks, with 16 active at a time. Model router OrcaRouter lists 2.8 trillion in total and about 104 billion active for each token. Moonshot's post gives no counts.
The context window, the amount of text the model can read at once, is 1 million tokens. A faster sibling, K3.1 Pro Highspeed, runs two to three times faster, according to Moonshot.
Benchmarks, as Moonshot reports them
Every score below comes from Moonshot and had not been checked independently at launch, The Next Web notes.
| Test | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 (coding agents) | 65.7% | 66.4% | 58.2% |
| BenchCAD, with tools | 92.7 | 96.2 | 95.9 |
| Hallu Bench Hard, with tools (lower is better) | 8.3% | 15.0% | 5.5% |
Moonshot ran Terminal-Bench 4.0 with the Claude Code harness, its footnotes say. On its own long-task coding test, Kimi Code Bench 3.0, K3.1 Pro scored 61.2, up from 55.95 for K3. The share of false claims on Hallu Bench Hard fell to 8.3% from 23.4% for K3. Moonshot warns those rates "do not represent overall error rates in daily use."
The cyber cut
The unusual part is what Moonshot took out. For the open-weight release, it says it removed cyber-related data from both mid-training and the reinforcement learning tasks the model practiced on.
It then published the cost. On xdev-bench-hard, an exploit-writing test, K3.1 Pro passed 81.0% of tasks. A version trained without the cut scored 90.5%. K3 scored 33.3%, and DeepSeek's V4.1 Flash scored 85.7%. A third-party run put OpenAI's GPT-5.6 Sol at 100%.
Moonshot says its model's cyber skills still trail the best closed models. It says it does not want open weights to push the frontier of attack capability, and plans to raise that ceiling step by step.
Price and access
Moonshot's post does not list prices. OrcaRouter, which resells model access, lists the API at $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for K3. Highspeed costs $4 and $20. Cached input costs $0.20.
| Model | Input $/M tokens | Output $/M tokens | Cache hit $/M tokens |
|---|---|---|---|
| Kimi K3 | 3.00 | 15.00 | 0.30 |
| Kimi K3.1 Pro | 2.00 | 10.00 | 0.20 |
| Kimi K3.1 Pro Highspeed | 4.00 | 20.00 | 0.20 |
The API model names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot offers zero data retention, and enterprise admins must switch the model on. OrcaRouter also reports fixed sampling: temperature 1.0 and top_p 0.95, with other values returning an error.
The politics around the weights
The release lands amid pressure on Moonshot. In July, the White House accused it of using banned Nvidia chips and copying an Anthropic model to build K3, without publishing evidence, The Next Web reports. Earlier this month, OpenAI tied a reasoning-extraction campaign to Moonshot.
China's own rules may matter more. A draft export-control list published on October 16 covers the weights of models with at least 2 trillion parameters. Reports cited K3 as above that line. Comments on the draft close November 15.
What this means for developers
If you run coding agents, K3.1 Pro is worth a trial on your own tasks. Moonshot's Terminal-Bench figure puts it level with Opus 5.5, at a lower list price than K3. Test it on your repositories before trusting a vendor score.
Plan around the fixed settings. Code that sets temperature or top_p will fail against this model, so check your client library and agent framework.
Do not build a self-hosting plan on November 4 yet. The weights are a promise, and Moonshot has not published a license. A model of about 3 trillion parameters also needs far more hardware than one workstation.
Security teams should note the trade. A model scoring 81% on an exploit test is still capable. Treat the open weights as cyber-capable when you assess risk, whatever the training cut.
Watch the K2 timeline. OrcaRouter reports that K2 models close to new users on October 31 and retire at the end of December 2026. Teams still on K2 should schedule a migration now.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Related on Tech AI Wire
- OpenAI links reasoning-extraction campaign to Moonshot AI
- China drafts export curbs on AI models and chip designs
Sources
- Introducing Kimi K3.1 Pro - Moonshot AI (Kimi)
- Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills - The Next Web
- Moonshot AI launches new flagship Kimi K3.1 Pro model with 3T parameters - SiliconANGLE
- Kimi K3.1 Pro: $2/$10 Launch and a Tougher Benchmark Sheet - OrcaRouter
Top comments (0)