Moonshot's Kimi K3 is a 2.8 trillion parameter open model with frontier level benchmarks
Moonshot AI's new open weight model posts GPQA and MathVision scores that read like a closed frontier lab's, and one arena leaderboard has it passing Claude on frontend code. Full weights land July 27.
A 2.8T parameter open model posted GPQA and MathVision scores that read like a closed frontier lab wrote them.
Moonshot AI released Kimi K3 this week, a 2.8 trillion parameter open weight model with a 1 million token context window and benchmark numbers that read like a closed frontier lab's. Full weights ship July 27.
What's actually in it
Kimi K3 runs on Moonshot's own Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture, a Mixture of Experts setup with 896 experts and 16 activated per token, MXFP4 weights with MXFP8 activations, and native vision. Moonshot claims roughly 2.5x the scaling efficiency of its predecessor, Kimi K2.
The benchmark numbers Moonshot published
DeepSWE: 67.5. Terminal Bench 2.1: 88.3. GPQA Diamond: 93.5. MMMU Pro with python: 83.4. MathVision with python: 97.8. These are Moonshot's own evaluation suite, not independently reproduced yet, but they are specific, checkable numbers rather than marketing language.
The claim that actually went viral: beating Claude on a coding leaderboard
Separately, the account behind a frontend code arena leaderboard posted that Kimi K3 now ranks first with 1,679 points, ahead of Claude Fable 5. That is one leaderboard, one moment in time, run by a third party, not part of Moonshot's own benchmark suite. It is also most of the reason this story is spreading past the usual open model release crowd.
Why a build studio cares
We pick a model per client build, not once for the year. An open weight model at this parameter count, with a published and checkable benchmark suite, is a real option now for cost sensitive or self hosted work, sitting alongside closed frontier APIs like Claude rather than replacing them outright. July 27, when the full weights ship, is the date worth marking, not the leaderboard screenshot from this week.
Next step: read Moonshot's own announcement for the full benchmark suite and license terms, and see the arena leaderboard post that started the Claude comparison. If you are weighing an open weight model against a closed API for a build, write to us at hello@gattyworks.com.