Kimi K3 is neck and neck with Fable5, with a combined accuracy rate of 93%

According to Vision One monitoring, the AI reasoning platform Fireworks used the same set of Agent scaffolding to test Kimi K3 and Fable 5, covering code fixes, long-range terminal operations, algorithms, multi-language programming, and legal tasks. Overall, the two were evenly matched, but excelled in different directions. K3 was stronger in security, cryptography, and long-term terminal operations, while Fable 5 excelled in multi-language programming, web development, and data visualization.

Fireworks then conducted a "God's-Eye-View Routing": each task ran two models simultaneously, and afterwards the correct and more cost-effective answer was selected. The combined accuracy reached 93%, higher than any single model. K3 handled 72% to 96% of the tasks, with Fable 5 only dealing with a few long-tail challenges.

K3 had lower costs in all five task categories. The long-term tasks were up to 50 times cheaper than Fable 5. However, the 93% is just a simulation result from the God's-Eye-View and not the actual performance results of Fireworks' routing decisions.

K3 will go live on Fireworks after its open-source release on July 27.