Kimi launches K3.1 Pro with one million token context window

Kimi has introduced K3.1 Pro, a new open-weight AI model designed for long-running work in coding, 3D design, and finance. In an official blog post, Kimi announces that the model has a context window of up to one million tokens and will be available through its apps, workplace tools, coding product, and API.

The company also released K3.1 Pro Highspeed, which it says responds two to three times faster than the standard model. Full model weights are scheduled for release by November 4, 2026.

K3.1 Pro is based on the earlier Kimi K3 model but uses more post-training compute, according to Kimi. The company says it improved the system through reinforcement learning, a training method that rewards successful behavior across tasks.

Focus on practical work

Kimi positions the model for complex, multi-step assignments. These include software development, financial document analysis, building spreadsheets, web research, and creating visual projects.

• Coding: Kimi reports a 65.7 percent score on Terminal-Bench 4.0, a benchmark for coding agents working in terminal environments.
• 3D creation: The model can generate CAD code and scripts for tools such as Blender, Kimi says. It also supports a new online 3D Make feature for previewing, editing, and exporting models.
• Finance: Kimi says the model improved its performance on financial modeling, research, and business simulation tests. It also reports a lower hallucination rate on its internal factual-accuracy benchmark.

The company says it applied restrictions to cyber-related training data for the open-weight release. In internal testing, Kimi reports that these measures reduced performance on vulnerability exploitation tasks compared with an unrestricted checkpoint, while preserving broader coding capabilities. Kimi says the model remains behind leading proprietary systems on offensive cybersecurity benchmarks.