Moonshot AI, the Beijing-based artificial intelligence startup, has introduced Kimi K3.1 Pro as the opening model in its K3.1 series. The company positions the system as its most capable open-weight offering in the roughly three-trillion-parameter class to date.
The release builds directly on the earlier Kimi K3 through substantially increased post-training computation, including large-scale reinforcement learning applied across hundreds of domains.
The resulting model retains a mixture-of-experts architecture with about 2.8 trillion total parameters and roughly 104 billion active per token, while adding a one-million-token context window and support for text plus image (and, in some descriptions, video) inputs.
A companion Highspeed variant is also available and is said to generate responses two to three times faster than the standard version.
Both are live now on Kimi’s website, mobile apps, Kimi Work, Kimi Code, and the company’s API, with full model weights promised for public release by November 4, 2026.
Pricing has been reduced relative to the prior generation.
The update focuses on stronger coding and agentic performance.
On Terminal-Bench 4.0, a test of coding agents handling realistic, resource-constrained terminal tasks with strict verification, the company reports a score of 65.7 percent for Kimi K3.1 Pro.
That figure places it ahead of OpenAI’s GPT-6 Astra and only narrowly behind Anthropic’s Claude Opus 5.5 in the company’s published comparisons.
An internal long-horizon coding evaluation likewise shows gains over the predecessor.
The model is described as capable of handling full-project work in game engines such as Unity, Unreal, and Godot, as well as generating and refining interactive web experiences.
In one internal demonstration it iteratively improved a complex human-walking dynamics simulator built on an open anatomical model, diagnosing issues and correcting underlying simulation errors rather than merely adjusting rewards.
Performance indicators also point to aptitude in three-dimensional design and finance.
On BenchCAD, which assesses reconstruction of three-dimensional objects from multi-view images through generation of CAD code, the model scores 92.7 percent when allowed tool use—approaching the results of the leading proprietary systems.
It is particularly noted for parametric modeling of mechanical parts, architectural forms, and configurable products, and it can drive tools such as Blender and CAD software to edit geometry, materials, and scenes.
In financial workflows the company highlights state-of-the-art results on an OfficeQA-style evaluation involving retrieval and multi-step numerical reasoning over real documents, plus measurable gains on modeling, research, and simulated business-operation tasks.
Targeted work on factual accuracy reduced claim-level hallucination rates on an internal hard benchmark from the prior generation’s level.
Kimi K3, released over the summer, had already drawn wide attention by posting competitive results that raised questions about the durability of U.S. leadership in frontier artificial intelligence and about whether China’s open-weight offerings were closing—or had already closed—the performance gap.
Independent evaluations of the new model place it fourth overall on the Artificial Analysis Intelligence Index shortly after launch, with the largest gains appearing in agentic tasks; once the weights appear it is expected to become the highest-scoring openly available model on that index and to narrow the remaining distance to the top proprietary systems.
The company also applied data-filtering measures during training to restrain certain vulnerability-research capabilities ahead of the open-weight release, trading some specialized performance for reduced misuse risk while still advancing general coding ability.
