Xiaomi live-streams MiMo-V2.6 RL training with public dashboard

After six months of quiet focus on how far reinforcement learning can scale, Xiaomi's MiMo team is live-streaming the entire RL training of MiMo-V2.6 and has launched a real-time training dashboard at mimo.xiaomi.com/rl/ that anyone can watch online — a move rarely seen with open-source models.

Confirmed

The MiMo team (@LuoFuli) announced that MiMo-V2.6 is currently in RL training, consuming roughly 2 billion tokens per step with multi-environment agentic RL; compute scale is one of the key scaling directions.
Model specs: MiMo V2.6 Pro has 1T total parameters with 42B activated; MiMo V2.6 Flash has 309B total parameters with 15B activated.
The stream discloses details including per-batch data and harnesses, all fully public.
Just one day into training, @zainhas observed that the deepswe benchmark had already reached about 60% progress.

Why it matters

Large-model RL training typically happens behind closed doors; broadcasting the training process in real time — including batch-level data and harnesses — is rare in the industry. Posters such as @skeole noted that many Reddit users were pleasantly surprised, with some saying that simply "watching it train" is a novel experience in itself.
The open broadcast offers a rare first-hand window into the actual progress and dynamics of RL scaling, and gives the community a reference point for evaluating how the next version of Xiaomi's MiMo series performs.

2026-09-17 ~ 2026-09-17 · 5 related posts

Primary sources

MiMo-V2.6 mid-RL run: ~2B tokens per step, multi-harness agentic RL, details to be open-sourced — bodonoghue85 · 2026-09-17 03:45 (UTC+08:00)
Xiaomi opens a live training dashboard for MiMo 2.6 RL runs — skeole · 2026-09-17 04:49 (UTC+08:00)
Mimo livestreams its RL training run for 1T-param V2.6 Pro with full internal metrics — SonglinYang4 · 2026-09-17 04:21 (UTC+08:00)

[source] MiMo-V2.6 mid-RL run: ~2B tokens per step, multi-harness agentic RL, details to be open-sourced — bodonoghue85 · 2026-09-17
Xiaomi livestreams MiMo-v2.6 RL training run, hits ~60% of deepswe in one day — zainhas · 2026-09-17
[source] Mimo livestreams its RL training run for 1T-param V2.6 Pro with full internal metrics — SonglinYang4 · 2026-09-17
Xiaomi opens a live training dashboard for MiMo 2.6 RL runs — Not-Post-Malone · 2026-09-17
[source] Xiaomi opens a live training dashboard for MiMo 2.6 RL runs — skeole · 2026-09-17