Moonshot AI has introduced Kimi K3.1 Pro, the first model in its new K3.1 series. The model features a 1 million token context window and is based on a 2.8 trillion parameter Mixture-of-Experts architecture, with 104 billion parameters active per token. It supports text, image, and video inputs.
The company offers two versions: K3.1 Pro and K3.1 Pro Highspeed, the latter being two to three times faster. Pricing for K3.1 Pro is $2 per million input tokens and $10 per million output tokens, with cached input costing $0.20. The Highspeed version is priced at double these rates.
Moonshot intentionally limited the model's cybersecurity capabilities by cleaning training data to prevent the exploitation of vulnerabilities. Despite this, the model achieved an 81% success rate on the internal xdev-bench-hard benchmark, up from 33% in the previous K3 version.
The model is available via API and Kimi services. Moonshot plans to release the full weights by November 4. Due to the architecture's size, which previously required clusters of at least 64 accelerators, the release is primarily aimed at cloud providers and large enterprises.
In performance benchmarks, K3.1 Pro shows strength in coding and agentic tasks, often matching or exceeding Claude Opus 5.5. However, it lags behind in complex reasoning tasks without tools. The model also demonstrated improved factual accuracy, reducing false claims in internal tests to 8.3%.
