Kimi K3 releases its weights and shakes up the frontier-model race

Moonshot AI has released the weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that activates 104 billion parameters at each step. The release turns a previously hosted-only system into a downloadable model and creates a new point of comparison with closed models from OpenAI and Anthropic.

What the release includes

Kimi K3 accepts text and images, offers a one-million-token context window and activates 16 of its 896 experts per token. Fireworks AI has made it available from day one for inference and training, with US-hosted endpoints and zero data retention.

Performance and availability

Ollama’s official model page lists cloud access at $3 per million input tokens and $15 per million output tokens. Fireworks publishes results across 663 coding tasks in which K3 approaches Claude Opus 5 at a lower cost per task. These are provider figures and should be tested on real workloads.

Confirmed availability includes weights on GitHub and Hugging Face, inference and training on Fireworks, and access through Ollama Cloud. At the time of this update, Nexus AI could not find an official Cursor announcement adding Kimi K3 to its model selector, so that integration is not presented as a fact.

What still needs to be tested

• Sustained performance beyond published benchmarks.
• Exact licence terms for products operating at very large scale.
• Cost and latency when using the full context window.
• Security of fine-tunes and derived models.