Moonshot AI Releases Kimi K3, the Largest Open-Weight Model Ever at 2.8 Trillion Parameters
Kimi K3's open weights landed a day early on July 26, and independent benchmarks put it neck-and-neck with Anthropic and OpenAI's proprietary frontier models.
Moonshot AI published the full open weights for Kimi K3 on July 26 at roughly 7:30 PM EDT, a day ahead of its planned July 27 target. At 2.8 trillion total parameters, the company calls it the largest open-weight model ever released — and early benchmarks show it trading blows with the best proprietary systems from Anthropic and OpenAI.
K3 is a mixture-of-experts model with 896 experts, of which only 16 activate per token — roughly 50 billion active parameters at inference time. That keeps compute costs manageable despite the model’s total size. It ships with a 1-million-token context window and native vision support, and the full weights download runs to 1.4TB, released under a Modified MIT license. Together AI and Modal both confirmed day-zero hosted inference access, so developers don’t need to wrangle a 1.4TB download just to try it.
The release timing isn’t an accident. It lands just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, and it marks a real comeback attempt for Moonshot. The company’s Kimi assistant ranked third in monthly active users in China before DeepSeek’s R1 release in January 2025 scrambled the entire domestic AI market — Kimi slid as low as seventh. Moonshot’s answer was to go all-in on open weights, starting with Kimi K2 in July 2025, then K2.5 in January 2026, and now K3.
That bet is paying off in relevance if nothing else. A 2.8-trillion-parameter open model that benchmarks close to frontier proprietary systems changes the calculus for any team currently paying API rates to Anthropic or OpenAI for tasks that don’t need the absolute best model — self-hosting or renting inference on a model this capable, for free as far as licensing goes, is now a real option. Independent testers have also flagged a 51% hallucination rate on certain evaluation sets, so K3 is not a drop-in replacement for every use case. Treat the size and the benchmark scores as impressive, and the reliability numbers as a genuine caveat, not a footnote.
For infrastructure teams, K3’s 896-expert MoE design is also a signal of where the open-weight frontier is heading: fewer dense giants, more sparse models that need less active compute per token but more total memory to host. Expect inference providers to compete hard on price for K3 access over the next few weeks.