Update: Kimi K3's 2.8T open weights drop — need 64+ Blackwell GPUs
Moonshot's 00:00 UTC window has arrived. Before you pull 1.4TB: ~50B active of 2.8T, reportedly tops frontend code, license still murky, hallucinations up.
Copy markdown
The 00:00 UTC drop window is here
Moonshot set K3's open weights for 00:00 UTC on July 27 — that hour has now arrived, ending weeks of "coming Monday." The full 2.8T-parameter model is due to land on Hugging Face at native MXFP4 precision.
A datacenter download, not a laptop one
Moonshot recommends 64+ accelerators in a single high-bandwidth domain, with native support only on NVIDIA Blackwell or AMD MI400. The checkpoint runs ~1.4TB at MXFP4 (about 1.5TB at BF16) before any KV cache — self-hosting is a cluster job.
Sparse by design: ~50B active of 2.8T
K3 is a Mixture-of-Experts with 896 experts and 16 firing per token, so only ~50B parameters are active per step. Inference cost tracks a mid-size model even though the checkpoint is frontier-scale.
Reportedly #1 on frontend code
K3 is reported to top Arena.ai's Frontend Code Arena at 1,679 points with a 76% blind win rate — the case for tolerating the hosting pain if you ship UI-heavy apps.
The catches: hallucinations up, license unclear
Independent testing reportedly puts K3's hallucination rate at 51%, up 12 points from K2.6's 39% — verify outputs in agent loops. And the license text still wasn't published alongside the weights, leaving commercial terms ambiguous.