Kimi K3’s full weights are here, but they’re ‘open’ with a caveat: What enterprises should know
Chinese AI startup Moonshot AI, creators of the increasingly famous “Kimi K” family of powerful, open AI models, today released the full weights for its largest and most performant version yet — Kimi K3. But enterprises evaluating the model should read the attached custom Kimi K3 usage license as carefully as the benchmark charts. VentureBeat previously covered Kimi K3 when it debuted through Moonshot’s hosted API earlier this month, including its 2.8 trillion-parameter architecture, one million-token context window and frontier benchmark performance. Today’s announcement completes that rollout with the release of the full model weights, a 47-page technical report documenting its training innovations and obstacles, and much of the infrastructure required to run the model independently. That includes the complete 2.8 trillion-parameter Mixture-of-Experts model, inference infrastructure, optimized attention kernels, MoE communication libraries and deployment components aimed at researchers and enterprise developers who want to self-host the system rather than consume it exclusively through an API. Moonshot is also releasing implementation support for ecosystems including vLLM and SGLang, alongside a technical report detailing the architectural innovations …









