Moonshot AI’s Kimi K3 Tops Fable 5 on Key Benchmarks

Moonshot AI released Kimi K3 on July 16. The model has 2.8 trillion parameters and will publish full model weights under a modified MIT license on July 27.

K3 uses a mixture-of-experts design with 896 expert subnetworks and activates only a subset for each task. The model supports a one-million-token context window, native image and video understanding, and continuous reasoning. Moonshot highlighted two techniques called Kimi Delta Attention and Attention Residuals, which it reports deliver about 2.5 times the scaling efficiency of its K2 predecessor.

Kimi Delta Attention speeds decoding on long contexts; Moonshot reports up to 6.3x faster performance at million-token contexts. Attention Residuals routes information selectively across layers and is reported to add roughly 25% training efficiency at under 2% additional compute cost.

On writing benchmarks, K3 scored 2,840 on a Writing Elo ranking, ahead of Fable 5 at 2,760. On a frontend code leaderboard, K3 reached an Elo of 1,679 versus Fable 5 at 1,631 and placed first in multiple frontend domains. In head-to-head web engineering tests, K3 won seven of eight tasks, including refactoring and debugging matchups. A composite index of nine evaluations placed K3 at 57 out of 100, with Fable 5 at 60 and GPT-5.6 Sol at 59.

Moonshot set K3 pricing at $3 per million input tokens and $15 per million output tokens. Internal per-task cost estimates across a nine-benchmark suite show K3 at about $0.94 per task, compared with $1.04 for GPT-5.6 Sol and $1.80 for an alternative model. Moonshot is offering free trials on its site but reports heavy traffic; paid subscriptions and API access are recommended for uninterrupted use.

The company will release the full model weights on July 27 under the modified MIT license, which will let enterprises run or fine-tune K3 locally. Moonshot notes that current commercial GPUs cannot host the full model without substantial resources.

Model documentation lists Nvidia H200 accelerators and a “GPGPU from an alternative vendor” among training hardware. At the Davos forum earlier this year, Moonshot AI president Yutong Zhang told attendees, “We knew we didn’t have the luxury to simply scale up compute… That forced us to focus on fundamental research and efficiency.”

K3 shows higher error rates on some measures. Its hallucination rate on an AA-Omniscience benchmark rose from 39% in the K2.6 predecessor to 51% in K3. Moonshot’s documentation also warns the model can be “excessively proactive,” and it advises users and developers to stress-test outputs and monitor behavior in extended autonomous tasks.