Kimi K3 Full Weights Arrive With 2.8T Parameters, Native Vision, and 1M Context

Kimi released the full K3 weights on July 27. The 2.8-trillion-parameter model combines native vision, up to a one-million-token context window, and agent-oriented capabilities.

What was released

Kimi introduced Kimi K3 on July 16, 2026, and released the full weights on July 27. Its official help center describes a 2.8-trillion-parameter multimodal model with native visual understanding and up to a one-million-token context window. K3 is also available through Kimi's web and mobile apps, Work, Code, and API.

Why it stands out

  • Releasing weights for a model at this scale expands research and self-hosting options.
  • The design targets work that mixes large codebases, long documents, and visual input in one context.
  • Kimi positions coding, long-running agents, and parallel Agent Swarm tasks as primary uses.

What to verify before deployment

A 2.8-trillion total parameter count does not mean convenient local inference. Teams still need to verify active parameters, quantization, GPU count and memory, and serving-framework support. Kimi's comparative benchmark claims should also be reproduced with the same harness and representative internal workloads.

Official sources