The Mach-Mind-4-Flash technical report introduces a 35B-parameter Mixture-of-Experts (MoE) agentic model with only 3B activated parameters. Through post-training optimization alone, the model achieves performance on par with or surpassing 100B-parameter-class models without scaling pre-training compute.
The pipeline includes three key stages: a unified RL/OPD infrastructure delivering a 17% training speedup; Multi-Teacher On-Policy Distillation (MOPD) to fuse domain-specific experts; and Hybrid Median-length Policy Optimization (HMPO), which compresses reasoning chains by 19--46% with minimal accuracy loss. The model scores 92.70 on AIME'26, 82.82 on IFBench, and leads or matches models with 10--30x its activated size.
By introducing scalable agentic interaction environments for reinforcement learning, the authors demonstrate significant performance gains on real-world application tasks at a fraction of the inference cost.