Kimi K3.1 Pro Becomes China's Most Powerful AI Model Despite Cybersecurity Training Restrictions
Chinese AI developer Moonshot AI has officially introduced Kimi K3.1 Pro, a new large language model that has emerged as China's highest ranked AI system on the Artificial Analysis Intelligence Index. The model achieved a score of 55, placing fourth globally at launch and outperforming several leading proprietary models from OpenAI and Anthropic. Remarkably, Moonshot achieved these results despite intentionally removing cybersecurity training data to limit the model's potential for offensive cyber operations before its planned open weight release.
According to Moonshot AI's official technical report, K3.1 Pro builds upon the previous Kimi K3 architecture, retaining 2.8 trillion total parameters and activating approximately 104 billion parameters per token through its Mixture of Experts design. The model also supports a native context window of 1 million tokens and uses substantially expanded reinforcement learning to improve coding, 3D modeling, financial analysis and autonomous agent workflows.
The performance improvements are particularly notable in complex programming tasks. Artificial Analysis reports that K3.1 Pro improved its Intelligence Index score by 11 points compared with K3, while achieving substantial gains in agentic coding and task automation.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published @Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, 44). It trails only Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index. Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3. Key results: ➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%. ➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%. ➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower. Additional model details: ➤ Context window: 1M tokens. ➤ Size: 2.8T total, 104B active parameters ➤ Input modalities: Text and image ➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch ➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates. Congratulations to the team at @Kimi_Moonshot on the release!
— Artificial Analysis (@ArtificialAnlys) October 23, 2026
Kimi K3 vs Kimi K3.1 Pro Performance
| Benchmark | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Intelligence Index | 44 | 55 |
| Terminal Bench 4.0 | 13% | 67% |
| AutomationBench | 58% | 73% |
| Internal Cybersecurity Test | 33.3% | 81.0% |
| Context Window | 1M tokens | 1M tokens |
| API Input Cost / 1M Tokens | $3 | $2 |
| API Output Cost / 1M Tokens | $15 | $10 |
Intelligence, coding and automation results are from Artificial Analysis. Cybersecurity results are Moonshot's internal evaluations and have not been independently verified.
The cybersecurity restrictions represent one of the most unusual aspects of this release. Moonshot identified and removed cyber related examples from its intermediate training data and reinforcement learning environments to reduce the model's ability to discover and exploit software vulnerabilities. An unrestricted checkpoint achieved 90.5% on the company's internal xdev benchmark, compared with 81.0% for the final model. Although the restrictions reduced its offensive cybersecurity performance by 9.5 percentage points, K3.1 Pro still significantly exceeds its predecessor's 33.3% result.
Moonshot's approach reflects the growing security challenges surrounding open weight AI models. Once model weights are publicly distributed, developers can deploy and modify them independently, making centralized restrictions more difficult to enforce. The company believes advanced AI should remain accessible for legitimate security research without unnecessarily expanding offensive capabilities.
Beyond cybersecurity, K3.1 Pro demonstrates improved abilities in game development and 3D content creation, including generating projects using Unreal Engine, Unity, Godot and Blender. Moonshot also introduced K3.1 Pro Highspeed, offering approximately 2 to 3 times faster inference at a higher API price.
Both models are currently available through Kimi, Kimi Code and the Kimi API, with full model weights scheduled for release by November 4, 2026. At $2 per million input tokens and $10 per million output tokens, K3.1 Pro is also 33% cheaper than K3 through the official API.
The launch intensifies competition among Chinese AI developers, particularly as Alibaba continues advancing its Qwen3.8 Max models for complex reasoning, coding and autonomous workflows. If Moonshot delivers its promised open weights, K3.1 Pro would become the highest scoring open weight model on Artificial Analysis at the time of its launch evaluation, narrowing the performance gap with proprietary AI systems.
Kimi K3.1 Pro demonstrates how quickly Chinese AI developers are closing the performance gap with leading proprietary models. Its strongest advantage is not simply its enormous parameter count, but the improvements in autonomous programming, complex reasoning and practical development workflows, including applications directly relevant to game development and 3D content creation.
However, the cybersecurity results raise an important question about how much capability developers can realistically restrict through training data filtering. Despite removing targeted examples, K3.1 Pro still performs exceptionally well in Moonshot's internal security evaluations, suggesting that advanced general reasoning and coding skills can transfer into cybersecurity tasks.
The planned open weight release could make the model particularly attractive to enterprises and researchers seeking greater deployment control, although running a 2.8 trillion parameter model locally would require substantial computing infrastructure. Its long term impact will depend on real world performance, hardware efficiency and whether Moonshot can balance accessibility with responsible AI development.
Would you trust an open weight AI model like Kimi K3.1 Pro for coding and cybersecurity research, or do you believe increasingly powerful AI models should remain under tighter developer control?
