Kimi-K2.7-Code Outperforms Fable-5 in ML Engineering Tasks

ME News reports that on June 15 (UTC+8), according to monitoring by Beating, AI agent developer WecoAI released evaluation results of seven cutting-edge large models on autonomous scientific research tasks. In the Machine Learning Engineering task, Kimi-K2.7-Code, the newly open-sourced trillion-parameter model from Moonshot AI, outperformed all other leading large models tested, including Anthropic’s flagship model Fable-5. The evaluation used a cost-constrained protocol—rather than step-limited—meaning that under a fixed budget, models with lower per-use costs could perform more attempts and iterations. Overall, while Fable-5 dominated the test suite in Prompt Engineering and Algorithm Discovery tasks and claimed the overall championship, its performance in Machine Learning Engineering lagged even behind its predecessor, Claude 3 Opus. The evaluation team suggested that Fable-5’s weaker performance in this task may be due to its high API costs placing it at a disadvantage under cost constraints, or because the task triggered more stringent safety guardrails within the model. (Source: BlockBeats)