Xiaomi open-sources MiMo-V2.6 RL dataset and training code
On Sept 26–27, Xiaomi's MiMo team open-sourced the MiMo-V2.6-RL-oss dataset (used to train MiMo-V2.6 pro/flash) and the accompanying training code on HuggingFace, quickly trending and widely seen as a milestone for open-source RL environment data.
Confirmed
- The dataset XiaomiMiMo/MiMo-V2.6-RL-oss is Apache-2.0 licensed, sized 1K–10K rows, covering text, image and document modalities in parquet format.
- It includes 7,000 RL task environments across code, cybersecurity, general tool use, visual web dev and music, unified in Harbor format with prompts, agent configs, reward metadata and a training framework.
- Training code lives in XiaomiMiMo/verl, forked from the ByteDance Seed-initiated, community-maintained verl post-training framework.
- Community reception was positive: @iScienceLuvr noted the release was praised as "rare"; @HarveenChadha pointed out such RL environments typically cost hundreds to thousands of dollars each; adithyask (via @SergioPaniego) called it a major advance for open-source RL environment data.
Unconfirmed
- Dorialexander, who examined the release (relayed by @teortaxesTex), clarified that the core is actually 989 envs rather than the full 7,000 — a discrepancy awaiting official clarification.
Why it matters
- High-quality RL task environments are typically commercially priced assets; by open-sourcing them fully under Apache-2.0 along with metadata and training frameworks, Xiaomi provided the community with reusable RL post-training infrastructure and a reproducible path for its model training stack.
2026-09-26 ~ 2026-09-27 · 9 related posts
Primary sources
- [source] Xiaomi MiMo open-sources its RL environments and training code on a fork of verl — eliebakouch · 2026-09-26
- Xiaomi open-sources the RL environment and code used to train MiMo v2.6 pro/flash — zainhas · 2026-09-26
- Xiaomi MiMo open-sources MiMo-V2.6-RL-oss training dataset on Hugging Face under Apache-2.0 — iScienceLuvr · 2026-09-26
- [source] Xiaomi Releases MiMo-V2.6-RL Dataset, Trends on Hugging Face — XiaomiMiMo · 2026-09-26
- Xiaomi open-sources ~7K RL environments used to train MiMo, spanning code, cyber and web dev — SergioPaniego · 2026-09-26
- [source] Xiaomi open-sources MiMo RL environments repo, individual tasks sell for $100-1000 each — HarveenChadha · 2026-09-26
- Xiaomi open-sources ~7K RL environments, but only 989 were used to train the 9B MiMo model — teortaxesTex · 2026-09-26
2 near-duplicate retellings: burny_tech · SergioPaniego