Baseten details day-0 API support for Moonshot's 2.8T-parameter Kimi K3
philipkiely · x · 2026-07-27
Baseten says it had day-0 support for Kimi K3 and used early-access weights from Moonshot AI to bring the model up on its Model APIs.
- Kimi K3 is described as a 2.8T-parameter open frontier model with Kimi Delta Attention, Attention Residuals, sparse experts, and a new vision encoder.
- The team says K3’s weights are over 1.4TB in MXFP4, so they ran it on NVIDIA GB300 NVL72 systems.
- They worked with vLLM and SGLang on pre-release inference engine builds, validating support for KDA, AttnRes, tool calling, tokenization, and vision inputs.
- The post emphasizes the engineering required to serve a trillion-plus-parameter open model at launch time.
More from Infra
- Cloudflare adds Moonshot’s Kimi K3 to AI Gateway, but Workers AI lacks day-one support — michellechen · 2026-07-27
- Baseten adds Kimi K3 to its Model APIs on launch day — baseten · 2026-07-27
- MoonshotAI open-sources MoonEP for perfectly balanced expert parallelism — teortaxesTex · 2026-07-27
- Moonshot releases Kimi K3, a 2.8T MoE model with 1M context and 423 tok/s serving — ricklamers · 2026-07-27
- How to run a private local AI on many 16GB laptops with Ollama and Gemma 4 — minchoi · 2026-07-27
- Vercel adds Kimi K3 on US providers with zero-data-retention support — cramforce · 2026-07-27