Summary
On July 27, 2026, Moonshot AI released the full open weights for Kimi K3 (about 1.4TB at MXFP4) under a Modified MIT license, following the model's July 16 API launch. K3 is a 2.8-trillion-parameter sparse mixture-of-experts model that activates roughly 50B parameters per token and ships a 1-million-token context for long-horizon coding and agentic work. It ranked first on the Frontend Code Arena at 1,679, ahead of Fable 5, though independent testers reported a roughly 51% hallucination rate that Moonshot omitted from its benchmark charts.
What changed
Moonshot released Kimi K3's full open weights (~1.4TB, MXFP4) under a Modified MIT license on July 27, 2026, moving the model from API-only availability to self-hostable, and published a Frontend Code Arena result (first place, 1,679, ahead of Fable 5).
Why it matters
The weights release turns the largest open model in the 3-trillion-parameter class into something enterprises can self-host, and beating a frontier closed model on a public coding benchmark reinforces that near-frontier agentic coding is commoditizing into open weights. The independently reported ~51% hallucination rate that Moonshot left off its charts is a caution that benchmark-topping open weights still need independent evaluation before production use.
Evidence excerpt
Arena ranked K3 first in its Frontend Code evaluation at 1,679 points, ahead of Fable 5 ... open weights release is scheduled for July 27 under a Modified MIT license ... independent testing found a 51% hallucination rate that Moonshot AI omitted from its benchmark charts.