Moonshot AI releases Kimi K3, making the 2.8-trillion-parameter model the largest open-weight AI ever downloadable

12:00 AM · July 27, 2026

Moonshot AI released the full open weights of Kimi K3 at 00:00 UTC on July 27, 2026, making the 2.8-trillion-parameter mixture-of-experts model, built with what the company calls Kimi Delta Attention and a Stable LatentMoE architecture, freely downloadable despite activating only 16 of its 896 experts, or about 1.8 percent, on any given token. The model uses MXFP4 weight quantization, which shrinks the full download to roughly 1.4 terabytes rather than the 5.6 terabytes full-precision weights would require, cutting memory bandwidth needs by around four times and making self-hosting viable for mid-sized organizations, though most users are expected to rely on inference providers given the hardware involved. On Moonshot's own launch benchmarks, K3 ranks first on Program Bench, SWE Marathon, SpreadsheetBench 2, Automation Bench, and BrowseComp, and second on Terminal-Bench 2.1 behind GPT-5.6 Sol, while independent testing reported by outlets covering the release noted a roughly 51 percent hallucination rate, a reminder that raw benchmark wins do not necessarily translate into reliability. On the community-voted LMArena leaderboard, K3 holds the top spot in frontend code generation with a 76 percent win rate against other models. Demand for API access was reportedly strong enough that Moonshot briefly capped new subscriptions in the run-up to the open-weight release, and the launch reinforces China's position at the frontier of open-weight AI even as it works around U.S. compute export restrictions.