Moonshot AI has released Kimi K3.1 Pro, a major upgrade to its open-weight model line that lands within striking distance of the strongest proprietary models on coding benchmarks. The 2.8-trillion-parameter mixture-of-experts model activates 104 billion parameters per token and carries a 1-million-token context window, making it one of the largest open-weight releases ever announced.
According to figures published by Moonshot, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, a demanding test of real-world terminal coding agents. That puts it just 0.7 points behind Anthropic’s Claude Opus 5.5 (66.4%) and ahead of OpenAI’s GPT-6 Astra (58.2%), though these are the company’s own numbers and not yet independently verified. It also leads Moonshot’s chart on SWE-Marathon at 51.7. Scientific terminal tasks remain a weak spot, where it trails GPT-6 Astra by a wide margin.
The model is live now in the Kimi app, Kimi Work desktop app, the Kimi Code coding agent, and the Kimi API at $2 per million input tokens and $10 per million output tokens — roughly a third cheaper than K3. A companion Highspeed variant runs two to three times faster for latency-sensitive work.
Full open weights are promised by November 4, 2026. Moonshot says it deliberately stripped cyberattack-related training data from the open release, keeping it behind frontier closed models on offensive cyber tasks. Developers who want to run open models on their own hardware can use Ollama, and model weights are typically shared through Hugging Face. To see how today’s leading closed models compare, read our Claude vs Gemini breakdown.
Source: MarkTechPost
