Moonshot AI Unveils 2.8T-Parameter Kimi K3.1 Pro With 1M-Token Context

Moonshot AI has unveiled Kimi K3.1 Pro, a 2.8-trillion-parameter model aimed at long-horizon coding and agentic workloads.
The model uses a mixture-of-experts architecture with roughly 104 billion parameters active per token and supports a context window of up to 1 million tokens.
Moonshot claims Kimi K3.1 Pro outperforms Claude Opus 5.5 and GPT-6 Astra on selected long-horizon coding benchmarks.
The company also reports leading results on SWE Marathon and several agent-focused evaluations.
Its ambitions extend well beyond chat.
Moonshot says the model can work on game projects across Unreal Engine, Unity and Godot, generate parametric 3D models and improve reliability in financial-modeling tasks.
The company plans to release open weights by November 4.
If those capabilities hold up outside controlled benchmarks, Kimi K3.1 Pro could become a major open-weight competitor in coding and agentic AI.
The key question is independent validation.
Benchmark leadership does not automatically translate into reliable real-world execution, especially on long, multi-step projects where errors can compound over time.