Together AI launches Moonshot Kimi K3 on day-zero; 2.8T open-weight model with 1M context window

Moonshot AI released the open-source weights for Kimi K3 on July 26, 2026, with Together AI and Modal both offering day-zero hosted inference access the same day, signaling a shift in infrastructure priorities toward immediate deployment of frontier open-weight models.

Kimi K3 is the first open-source model to reach 2.8 trillion parameters, featuring a Mixture-of-Experts architecture with only 16 of 896 experts active per token (Stable LatentMoE), native multimodal vision, and a 1-million-token context window using Kimi Delta Attention (KDA) to enable long-horizon reasoning at linear rather than quadratic cost.

Together AI positioned K3 as a launch platform across its full inference stack: Serverless (dev iteration), Provisioned Throughput (reserved capacity with SLA), and Dedicated Model Inference (full control). The partnership includes day-zero post-training—developers can fine-tune K3 directly on Together using full-weight and LoRA methods without leaving the platform.

For infrastructure teams, the timing signals how competitive open-weight models have become: K3 benchmarks claim parity with Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on some evaluations, and day-zero availability in production platforms removes friction for adoption. This is not a research release—Cursor, Y Combinator portfolio companies, and Decagon have begun deploying K3 immediately through hosted inference APIs.