❯ Moonshot releases Kimi K3.1 Pro weights as the model leads open entries on Agent Arena

Developers can obtain the weightsMoonshot has released Kimi K3.1 Pro on Hugging Face, fulfilling its previously announced plan. The model accepts text and images and targets improvements in coding, 3D creation, financial research and complex agent tasks. Developers can download it for their own deployments or use the Kimi API.

Real tasks provide another view of capabilityKimi K3.1 Pro (Max) ranked first among open-weight models and 13th overall on Agent Arena, leading open models in work, coding and chat, IT Home reported on November 5. The evaluation covered more than 17,500 real long-horizon sessions, using task confirmations, corrections, complaints and command errors to estimate performance. Its net improvement rate rose from +0.9% for its predecessor to +5.1%.

Sparse architecture reduces long-context overheadThe mixture-of-experts model has about 2.8 trillion total parameters, activates roughly 104 billion per token and supports a context window of about one million tokens. Each computation uses a subset of parameters, while hosting still requires memory for the large weights. The model was post-trained from Kimi K3 using large-scale joint reinforcement learning, replacing some full-attention layers with sparse-attention layers for more efficient long-context processing.

Terminal results sit alongside deployment costsThe model card reports 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 in its comparison table. Kimi used the Claude Code harness for these practical terminal tasks. Obtaining weights gives companies control over serving and optimization, alongside the need for substantial memory, multiple GPUs and inference operations.

▮ SIGNALStronger open weights give companies control over deployment and serving, with hardware requirements, inference efficiency and operations jointly determining practical costs.