Inside MiniMax: RL Lead Details Infrastructure and Training for Agents at Scale
AI Engineer · youtube · 2026-07-31
Olive Song, Head of Reinforcement Learning at MiniMax, shared an in-depth look at the infrastructure and tech stack behind the company's open-weight models.
Key Insights & Engineering Practices:
- Value of Open Source: Believes that releasing weights allows builders to optimize, helping open models close the gap with proprietary ones faster than expected.
- Agent Training: Models are trained not only for agentic coding but also to build games; they acquire computer-use skills via RL in environments like OS World.
- Day Zero Inference Stack: Requiring the inference stack to be ready at launch means writing and tuning custom GPU kernels, mastering parallel kernel benchmarks, and threading optimization through KV cache handling and routing.
- Multimodal Pitfalls: Text and vision capabilities tend to collapse after training unless both modalities are trained jointly.
- Long-Horizon Tasks: The team successfully replicated and evaluated 12-hour runs to ensure model stability over complex, extended tasks.
More from coding & agent
- LLM Agent Autonomously Conducts CT Reconstruction Research, Matching SOTA with 969 Parameters — maier_ak · 2026-07-31
- Inkling-Small Available on Together AI for Cost-Effective Agentic Coding — togethercompute · 2026-07-31
- New Benchmark: 200+ Sokoban Rooms to Test Agent Planning Skills — generativist · 2026-07-31
- Alibaba's Qwen-UI-Agent: SOTA on Mobile Use, Beats GPT-4o — AlibabaTongyiLab · 2026-07-31
- LedgerMind Tackles Multimodal Agent Hallucination via Structured Evidence Ledgers — Enjun Du · 2026-07-31
- AI Tour Meeting: A Multi-Agent Framework for Group Travel Planning — Daisuke Kikuta · 2026-07-31