Kimi K3.1 Pro boosts AI agents and 3D code
Moonshot AI introduces Kimi K3.1 Pro, a large language model with 3 trillion parameters. The system targets code, agents, and 3D, with immediate availability and open weights announced for November 4.
Moonshot AI unveiled Kimi K3.1 Pro, a large language model designed to compete with Claude Opus 5.5 and GPT-6 Astra on code, analysis, and certain 3D creation uses. The Chinese company also highlights a one million token context, a capacity designed for AI agents and long codebases, as companies seek less a chatty chatbot and more a tool that truly executes. The model is available at launch via Moonshot AI's paid services, with a Highspeed version, and its weights are to be opened on November 4.
Un modèle géant, mais pas seulement un concours de chiffres
Kimi K3.1 Pro is based on the Kimi K3 architecture, launched in July, which totals 896 neural networks for 3 trillion parameters, of which only 16 are activated per request to limit inference costs. Moonshot AI also retains Kimi Delta Attention, a memory mechanism designed to reduce RAM usage during prompt processing. In short, the model grows on paper but tries not to overwhelm the infrastructure in the process.
The Pro version received enhanced post-training, with more computation, automated quality controls, and an improved tech stack. According to Moonshot AI, this work mainly serves agentic workflows tasks, that is, chains of actions where the model reasons, plans, then executes successive steps, rather than responding in a single well-phrased sentence.
Des résultats solides sur le code, la 3D et l’analyse financière
On the Terminal-Bench 4.0 benchmark, Kimi K3.1 Pro achieves 65.7% success, compared to 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, according to figures published by Moonshot AI. The model also surpasses its rivals on OfficeQA Pro v2, a test of complex financial document analysis. On BenchCAD, which evaluates code generation for building 3D objects, it scores 92.7%, less than four points from the best scores announced by the company.
Moonshot AI also claims the model can animate 3D structures, simulate human movements, and produce executable code for video games. These uses remain challenging for AI systems, as a simple code error can break everything. Here, the promise is therefore less about spectacular rendering than the ability to produce something usable in a technical environment.
La sécurité reste le point faible assumé
Cybersecurity performance lags behind Claude Opus 5.5 and GPT-6 Astra. Moonshot AI explains it deliberately removed data related to vulnerability analysis from the entire post-training set to reduce security risks. The message is clear: the company prefers to limit certain capabilities rather than open the door too wide. It is cautious, but this leaves a big gap in a strategic area for enterprise use.
The company plans to equip its future models with more advanced vulnerability analysis functions. At this stage, the hierarchy between power, security, and control remains in flux, which matters for technical teams who will have to integrate the model into often highly locked internal environments.
Key points
- 3 trillion parameters, with 16 networks activated at once.
- 1 million token context for long codebases.
- 65.7% on Terminal-Bench 4.0, according to Moonshot AI.
- 92.7% on BenchCAD, for 3D generation by code.
- Open weights announced for November 4.
- Moonshot AI compares it to Claude Opus 5.5 and GPT-6 Astra.
By the numbers
- 896 neural networks — Kimi K3 architecture, launched in July, according to Moonshot AI.
- 16 networks activated — per request, to reduce inference cost.
- 65.7% — score on Terminal-Bench 4.0, published by Moonshot AI.
- 92.7% — score on BenchCAD, 3D modeling benchmark by code.
- November 4 — announced date for open weight release.