[Kimi K3 and Fable 5 Nearly Tied, Combined Accuracy Reaches 93%] According to monitoring by Beating, the AI inference platform Fireworks tested Kimi K3 and Fable 5 using the same Agent framework, covering five types of tasks: code repair, long-term terminal operations, algorithms, multilingual programming, and legal tasks. The two models were overall nearly tied but excelled in different areas. K3 performed better in security, cryptography, and long-term terminal operations, while Fable 5 excelled in multilingual programming, web tasks, and data visualization. Fireworks conducted a 'God's-eye view routing' test: both models ran simultaneously for each task, and the correct and cheaper answer was selected afterward. The combined accuracy reached 93%, higher than any single model. K3 handled 72% to 96% of the tasks, while Fable 5 only dealt with a few long-tail challenges. K3 also had lower costs across all five task categories, with long-term terminal tasks being up to 50 times cheaper than Fable 5. However, the 93% accuracy is only a simulated result under the God's-eye view and does not represent the actual routing performance achieved by Fireworks. K3 will be open-sourced and launched on Fireworks on July 27. [Original Link]