Moonshot's newest model nearly matches the best closed AI at coding — and the most interesting part is what the company left out.
October 24, 2026 — Beijing-based Moonshot AI has launched Kimi K3.1 Pro, an open-weight AI model that, according to the company, nearly matches Anthropic's Claude Opus 5.5 on a coding-agent test — while deliberately dialling back its hacking ability before its weights go public.
Image: A glowing neural-network visualization above rows of servers in a dark data centre. Credit: AI-generated illustration / Red Eye Eagle News.
Editor's Highlight: A leading AI lab has openly admitted its model was getting better at hacking as training scaled — and shipped a weaker version on purpose, because open weights cannot be taken back.
Table of Contents
- What Moonshot Just Released
- The Deliberate Cyber Pullback
- How It Scores Against the Frontier
- The Price Cut
- Why Open Weights Change the Stakes
- What to Watch Before November 4
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Model | Kimi K3.1 Pro (Moonshot AI) |
| Launched | October 23, 2026 |
| Architecture | Mixture of experts: 2.8T total parameters, 104B active |
| Context window | 1 million tokens |
| Faster variant | Kimi K3.1 Pro Highspeed, 2–3 times faster |
| API price | $2 input / $10 output per million tokens |
| Open weights | Promised by November 4, 2026; licence unannounced |
| Cyber benchmark | 81.0% released model vs 90.5% unfiltered version |
What Moonshot Just Released
Kimi K3.1 Pro is the first model in a new K3.1 series, and Moonshot describes it as its most capable 3-trillion-parameter-class open-weight model to date. It is not a brand-new base model: it is a post-trained version of Kimi K3, given far more reinforcement learning across what the company calls hundreds of downstream domains.
The model keeps K3's scale — roughly 2.8 trillion total parameters with 104 billion active — and a 1-million-token context window, so it can take in very large documents or codebases in one request. It accepts text and image inputs, and is live now in the Kimi app, Kimi Work, Kimi Code and the Kimi API. Moonshot says the full model weights will be published by November 4.
The Deliberate Cyber Pullback
The most unusual part of the launch is what Moonshot says it removed. The company found the model's ability to find and exploit software flaws rising steadily as training scaled, so for the open-weight version it stripped cyber-related data from training, keeping general coding skills.
On Moonshot's in-house exploit benchmark, which reproduces recently disclosed bugs in software like Linux and Node.js, an unfiltered checkpoint solved 90.5% of tasks, while the released K3.1 Pro solved 81.0%. Moonshot puts K3 itself at just 33.3% on the same test. "We aim not to release open weights that would extend the absolute frontier of cyber capability," the company said. The results have not been independently verified.
Analysis Note: The 81.0% figure cuts both ways — lower than the unfiltered model, but still high, and once weights are public, anyone can fine-tune the model back in the other direction.
How It Scores Against the Frontier
All benchmark figures below are Moonshot's own and have not been independently checked. One independent datapoint exists: Artificial Analysis puts K3.1 Pro fourth on its Intelligence Index at 55, behind only three closed models and 11 points above K3.
| Benchmark | Kimi K3.1 Pro | Closest rival |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5: 66.4% |
| BenchCAD, with tools | 92.7% | Claude Opus 5.5: 96.2% |
| OfficeQA Pro v2 | 64.4% | Claimed best score |
| Hallucination test | 8.3% | K3: 23.4% |
On Terminal-Bench 4.0, which tests coding agents on real terminal tasks, K3.1 Pro's 65.7% lands just behind Claude Opus 5.5's 66.4% and ahead of OpenAI's GPT-6 Astra at 58.2%. Moonshot took rivals' scores from their published reports or official leaderboards.
The Price Cut
K3.1 Pro is cheaper than the model it follows. It costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 — down from $3 and $15 for K3, a cut of about a third. Highspeed costs $4 and $20.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Kimi K3.1 Pro | $2 | $10 |
| Kimi K3.1 Pro Highspeed | $4 | $20 |
| Kimi K3 (previous) | $3 | $15 |
| Claude Opus 5.5 | $4 | $20 |
At those list prices, the standard K3.1 Pro costs half as much as Claude Opus 5.5. One practical limit: K3.1 Pro only works in thinking mode, according to Kimi's model documentation.
Why Open Weights Change the Stakes
Open-weight releases differ from hosted models: with a closed service, the developer can restrict access and update safeguards, but once weight files spread, it has far less control over who runs the model or what they do with it.
That is the backdrop to Moonshot's cyber filtering. Governments, technology companies and security researchers warn AI could make sophisticated hacking easier and cheaper, as models that write code and run tools may help users find software weaknesses and automate parts of attacks.
China proposed export controls last week that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools. The draft measures, open for public comment, cite risks to critical infrastructure, defence-industrial and core government systems. Moonshot argues the other side: open weights can help more organisations inspect and secure their own software, including older systems.
What to Watch Before November 4
Three dates matter now. First, November 4, when the weights are due — watch for the licence terms, still unpublished. Second, watch for independent verification: the benchmark and cyber-filtering claims can only be tested once researchers can run the weights.
Bottom Line
Kimi K3.1 Pro is the closest an open-weight model has come to the closed frontier on coding work, at half the API price of its nearest rival — and Moonshot is unusually open about the trade it made to get there safely. The real verdict arrives on November 4, when the weights, licence and cyber claims face independent testing at once.
FAQ
What is Kimi K3.1 Pro?
It is Moonshot AI's newest large language model, a post-trained version of Kimi K3 with 2.8 trillion total parameters, 104 billion active parameters, and a 1-million-token context window, available through Kimi's app, Work, Code and API.
When will the open weights be released?
Moonshot has promised the full weights by November 4, 2026. Until then, only its apps and API can run the model, and the licence is unannounced.
Why did Moonshot weaken the model's hacking ability?
The company says its hacking ability kept rising as training scaled. Because open weights cannot be recalled once distributed, Moonshot removed cyber-related data from training.
How does it compare with Claude Opus 5.5 and GPT-6 Astra?
On Moonshot's own figures, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. Artificial Analysis ranks it fourth overall at 55, behind only three closed models.
Can I run Kimi K3.1 Pro on my own computer?
Not yet, and not on ordinary hardware even after release: at around 3 trillion parameters, it needs data-centre hardware.
Sources
- Reuters (via Superhits 97.9) — https://superhits979.com/2026/10/24/moonshot-curbs-hacking-ability-in-new-kimi-model-as-fears-grow-over-ai-powered-cyberattacks/
- MadRobot — https://madrobot.blog/2026/10/23/kimi-k3-1-pro-moonshot-open-weight-nearly-matches-claude-opus-5-5-coding/
- Skimwire — https://www.skimwire.com/moonshot-kimi-k3-1-pro-open-weight-model/
- AI-generated illustration — ~/workspace/goals/red-eye-eagle-news-publishing-desk/hidden_files/hero-moonshot-kimi-k3-1-pro-2026-10-24-morning.jpg
Join the conversation
Comments
Post a Comment