Moonshot releases Kimi K3, a 2.8T MoE model with a 1M-token context window
burny_tech · x · 2026-07-28
Kimi K3 debuts as a 2.8T MoE model with 1M-token context
Moonshot says Kimi K3 is its most capable model yet: a 2.8T MoE model with native visual understanding and a 1M-token context window. The company claims a new architecture that delivers 2.5x intelligence per unit of compute, framing the release as an efficiency leap rather than a pure parameter-scale race.
Alongside the model weights and technical report, Moonshot is also opening parts of the stack behind K3, including high-performance attention kernels, an MoE communication library, and infrastructure for running agent environments at scale. A comment in the thread notes the model uses aggressive sparsity: around 2% active experts, or 16 out of 896 experts total.
Related event: Moonshot Releases Open-Weight Kimi K3 Model(138 posts)→
More from Infra
- Inference.net pitches a gateway flow that mirrors prod traffic to Kimi K3 before switching — MatthewBerman · 2026-07-28
- AI is likely to control quantum computers first, then use them for narrow science tasks — imjustnewatai · 2026-07-28
- Moonshot’s Kimi K3 lands in Japan with 2.8T open weights and $3/$13 pricing — DavidBennett__ · 2026-07-28
- A Kimi-k3 joke contrasts a $1,908 annual plan with $1.0908M to run it at home — HarveenChadha · 2026-07-28
- Calibrated Qwen3.6-27B quantization tests weight groups before compressing them — enginetown · 2026-07-28
- Frozen 12B system reuses verified memory at zero tokens and 6,000,000-token context — Corbenci · 2026-07-28