Li Auto just dropped the details on how it built Mach-Mind-4-Flash — and the answer involves cramming over a dozen specialized AI experts into a single model.
The company trained separate models for math, code, writing, safety, search, and tool calling. Then it merged them using a technique called MOPD (multi-teacher online policy distillation). Each question gets guidance from the corresponding expert, so the different skills don't trip each other up during mixed training.
The report also spells out how they trained the agent. The model doesn't just learn tool names and call formats — it actually reads files, edits code, runs tests, and calls tools inside an isolated sandbox. When an operation fails, it sees the error and environment changes, then keeps trying. The tool training covers over 190 stateful domains and more than 3,500 tool interfaces; a single programming trajectory can stretch to 300 rounds.
But this fusion isn't entirely lossless. Instruction following mostly holds up — ClawBench and ClawEval even beat the standalone agent expert. Yet SWE-bench Verified dropped from 73.80 to 71.10. The team admits that highly specialized abilities like long-range coding still suffer during merging.
The report also reveals how they reduce the model's tendency to overthink. HMPO dynamically sets a reasoning budget based on the median length of correct answers, only rewarding correct responses that are shorter. It cuts generation length by 19% to 46% with at most a 0.7 percentage point accuracy drop. For now, this method only applies to single-turn reasoning, not yet extended to multi-turn agent tasks.