< Back to all clusters
[TECHNOLOGY] · China · 34 sources

started · updated

Moonshot AI unveils Kimi K3.1 Pro with 3 trillion parameters

Moonshot AI has launched Kimi K3.1 Pro, a 3-trillion-parameter open-weight large language model featuring a 1-million-token context window. The release includes a Highspeed variant that offers two to three times the output speed of the standard version.

The model demonstrates significant performance in several specialized areas. In coding tasks, it scored 65.7% on the Terminal-Bench 4.0, surpassing GPT-6 Astra. It also achieved a 92.7% score on the BenchCAD 3D modeling benchmark and a 64.4% score on the OfficeQA Pro v2 financial document test. Internal stress testing reported that factual hallucination rates dropped from 23.4% in the previous K3 model to 8.3%.

To address cybersecurity concerns, Moonshot AI intentionally limited the model's ability to generate exploits by removing specific cybersecurity and exploit-generation data from its training and reinforcement-learning pipelines. This resulted in a 9.5-point drop in success rates for vulnerability exploitation in internal tests, from 90.5% to 81%.

Kimi K3.1 Pro is currently available via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The company plans to release the full model weights by November 4, 2026.

Entities

Amazon AWS · Artificial Analysis · Beijing · Claude Opus 5.5 · GPT-6 Astra · GitHub · Hong Kong · KIMI · Kimi K3.1 Pro · Moonshot AI

Claims

What the coverage asserts, and how many sources carry each claim.

Sources