Somewhere in the AI arms race, "bigger" became a personality trait. Moonshot AI just took that trait to its logical extreme with Kimi K3, a 2.8-trillion-parameter open-weight model that makes every other "open" release this year look like a rounding error.

896 Experts, 16 Show Up to Work

Kimi K3 is officially the largest open-weight AI model released to date, and it got there without demanding a small country's power grid, thanks to an ultra-sparse Mixture-of-Experts design. Out of 896 total specialized expert networks packed into the model, only 16 activate for any given prompt — so you get trillion-scale capacity without trillion-scale compute bills every time someone asks it a question.

The model ships with a 1-million-token context window, native vision support, and always-on reasoning. Moonshot released it on July 16, with the full 1.4-terabyte weight set (compressed via MXFP4 quantization) landing July 27 — timed neatly to land just before the World Artificial Intelligence Conference in Shanghai.

Open Weights, Open Warfare

This matters because Kimi K3 isn't just big for bragging rights — in blind developer testing on the Frontend Code Arena benchmark, it ranked first, ahead of Anthropic's Claude Fable 5. That's a Chinese open-weight model beating a closed frontier release from a leading US lab on a real coding benchmark, not a cherry-picked internal eval.

The subtext nobody's saying out loud in the press releases: this is China demonstrating that US compute export restrictions haven't stopped it from fielding models that trade blows with Silicon Valley's best. Every time a lab claims a compute moat is the thing keeping frontier AI contained to a handful of countries, releases like this quietly argue otherwise.

Somewhere, a GPU cluster the size of a small stadium is very tired, and it's not even the American one this time.

Source: VentureBeat