Moonshot AI Kimi K2.5
Moonshot AI released Kimi K2.5 as an open-source, natively multimodal, agentic Mixture-of-Experts model: 1 trillion total parameters with 32B activated, a 256K-token context, and continual pretraining on ~15T mixed visual-and-text tokens atop Kimi-K2-Base. It accepts images and video, runs in both 'thinking' and 'instant' modes, can generate code from visual specifications, and features an 'Agent Swarm' mode coordinating many sub-agents in parallel. Weights and code are released under a Modified MIT License, positioning it to compete on coding and agentic benchmarks at low cost.
The model is genuinely open and available: the official MoonshotAI GitHub repo and Hugging Face model card publish the weights, architecture, and benchmark tables comparing it to GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro. Its existence is independently corroborated by competitors — Z.ai's own GLM-5 benchmark table lists Kimi K2.5 as a peer model.
Kimi K2.5 is a real, downloadable, trillion-parameter open model and a strong signal that open multimodal-agentic systems are now competitive on coding and agentic work at a fraction of closed-model pricing. Because the weights and a technical report are public, its capability claims are verifiable rather than purely promotional, which is exactly the kind of transparency executives should reward. The practical caveat is scale: 1T parameters (even at 32B active) is not trivial to self-host, so the realistic path for most organizations is a hosted endpoint, which reintroduces the data-governance questions open weights were supposed to solve. As with GLM-5, the headline economics are real on paper but depend on your inference setup and on whether the benchmarks match your actual tasks.
A genuinely open, multimodal, agentic model at trillion-parameter scale pressures closed-model pricing and gives teams an auditable alternative for coding and agent workloads.