Xiaomi MiMo-V2.6 scales RL to 3.7B tokens per step, targeting self-improvement
Dr_Singularity · x · 2026-09-23
- Xiaomi's MiMo-V2.6 is its biggest RL scaling effort yet, spanning coding, general reasoning, vision, automation, and cybersecurity.
- Each RL step processes 1,568 samples and 2.7–3.7 billion tokens, with context lengths up to 1M tokens.
- Instead of narrow benchmarks, training uses multiple complex environments and stronger agentic graders to improve long-horizon reasoning and push for more efficient solutions.
- Benchmark performance keeps rising throughout training, suggesting large-scale RL still has substantial headroom; Xiaomi explicitly frames this as scaling toward model self-improvement.
More from Models
- DeepMind launches AlphaGenome 2 with major performance leap in variant-effect prediction — anshulkundaje · 2026-11-11
- US data vendors accused of selling RL environments to Chinese military-linked firms — xeophon · 2026-11-11
- Theory: Typos Disrupt LLM Tokenization, Which May Explain Why AI Text Has a Distinct Smell — menhguin · 2026-11-11
- LMArena co-founder explains AI evaluation methodology and policy in new podcast — MattPerault · 2026-11-11
- Runway open-sources Praxis-1, turning video pretraining into real robot control — runwayml · 2026-11-11
- DeepMind launches AlphaGenome 2, spotting nearly 2x more gene-regulating mutations — GoogleDeepMind · 2026-11-11