Hacker News
Kimi K3.1 Pro and K3.1 Pro Highspeed
AIFirst seen 1,066 points655 comments
Moonshot AI's Kimi team has released Kimi K3.1 Pro, a large open-weight AI model, plus a faster version called Kimi K3.1 Pro Highspeed. The company says the new model is strong at coding, 3D design, and research tasks.
What happened
The Kimi team introduced Kimi K3.1 Pro, the first model in a new K3.1 series. It is its most capable open-weight model so far, and it can read up to 1 million tokens of text at once. They also released Kimi K3.1 Pro Highspeed, which runs 2 to 3 times faster for quicker replies. Both are available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The full model weights, the numbers that make up the model, will be released by November 4.
Key numbers
- Terminal-Bench 4.0 coding test: K3.1 Pro scores 65.7%. That beats GPT-6 Astra (58.2%) and is slightly behind Claude Opus 5.5 (66.4%).
- Kimi Code Bench 3.0 Long-horizon (the company's own test): K3.1 Pro scores 61.2%, up from K3's 55.95%.
- BenchCAD 3D test: K3.1 Pro scores 92.7% with tools. Claude Opus 5.5 scores 96.2% and GPT-6 Astra scores 95.9%.
- Highspeed version: 2 to 3 times the speed of K3.1 Pro.
- Context window: 1 million tokens.
What is new
Kimi K3.1 Pro is an improved version of Kimi K3. The team used much more training time on it, called post-training, with reinforcement learning. This is a method where the AI learns from trial and error, getting rewards for good results. The company says K3.1 Pro can build complete games and web pages, work with tools like Unity and Godot, and create 3D models from a text description, sketch, or photo. In one test, it built a simulated human that learned to walk for 11.2 seconds without falling.
Why it matters
The article shows that open-weight models, which anyone can download and study, are getting close to the best closed models from big companies. If the scores hold up, developers and students may get strong coding and design help for free or at lower cost. The company's own tests, however, are not independent checks, so outside testing will matter.
What to watch
- Whether the full model weights are released by November 4.
- Independent test results that compare K3.1 Pro with GPT-6 Astra and Claude Opus 5.5.
- How well the Highspeed version performs in real use.
- How developers use K3.1 Pro for games, 3D design, and coding projects.
Glossary
- Context window
- The amount of text an AI model can read and keep in mind at one time, measured in tokens.
- Terminal-Bench 4.0
- A test that checks how well AI coding agents complete tasks in a command-line terminal.
- BenchCAD
- A test that checks how well an AI can rebuild 3D objects from several pictures by writing CAD code.