Xiaomi live-streams MiMo-V2.6 RL training, burning over $1.08M in 36 hours

After nearly half a year of quiet research into how far RL can scale, Xiaomi's MiMo team is live-streaming the full RL training of MiMo-V2.6 via a public dashboard at mimo.xiaomi.com/rl/. Training cost topped $1.08M within 36 hours ($30K/hour); APPSO later reported combined costs for the Pro and Flash variants exceeding $1.16M at $30.9K/hour, with Pro consuming 25.1B tokens cumulatively. Spending, steps, reward curves and even single-GPU failures are all publicly visible.

Confirmed

The MiMo team (Luo Fuli, @LuoFuli) said each RL step uses roughly 2B tokens (1568 prompts × 16 rollouts) with multi-environment agentic RL, and that compute scale is a key scaling axis.
Two models train in parallel: MiMo V2.6 Pro (1T total, 42B active) and MiMo V2.6 Flash (309B total, 15B active).
The stream exposes per-batch data, harness details, metric curves, and even VRAM fluctuations in real time.
Cost figures per QbitAI (over $1.08M in 36 hours) and APPSO (over $1.16M combined, $30.9K/hour, 25.1B tokens for Pro).
@zainhas observed DeepSWE benchmark progress reached 60% within one day of training start.
@karminski3 estimated the two models are in the agentic RL post-training phase, evaluated on DeepSWE (Pro scoring 62, compared against the DeepSeek-v4.1 series), and inferred roughly 6000-8000 H100 GPUs from the $10/second burn rate.
@MaziyarPanahi noted one RL step consumes 2B tokens — equal to what their own Pro-subscription agent uses in 12 hours — and called on more labs to open up training.

Why it matters

Large-model RL training normally happens behind closed doors; real-time public training with batch-level data, costs and failures is extremely rare. @skeole and @zainhas noted Reddit users were delighted by the novelty of "watching it train."
The live stream offers a rare first-hand window into RL scaling dynamics and a reference point for evaluating Xiaomi's next MiMo release.

2026-09-17 ~ 2026-09-17 · 14 related posts

Primary sources

Xiaomi opens MiMo-V2.6 live RL training panel: $1.16M+ spent, rewards still climbing — APPSO · 2026-09-17 11:29 (UTC+08:00)
Xiaomi livestreams MiMo-V2.6 RL training, burning over $1.08M in 36 hours — 量子位 · 2026-09-17 08:56 (UTC+08:00)
Xiaomi's Livestreamed RL Training Burns ~$10/sec, Est. 6,000-8,000 H100s — karminski3 · 2026-09-17 06:31 (UTC+08:00)

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
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
Xiaomi opens a live training dashboard for MiMo 2.6 RL runs — skeole · 2026-09-17
[source] Xiaomi's Livestreamed RL Training Burns ~$10/sec, Est. 6,000-8,000 H100s — karminski3 · 2026-09-17
Xiaomi livestreams model training burning ~$10/second, mimo models in early Agentic RL stage — karminski3 · 2026-09-17
Mimo livestreams full RL training of V2.6 Pro (1T, 42B active) with internal metrics — mishig25 · 2026-09-17
Xiaomi livestreams MiMo v2.6 RL training runs in rare transparency move — burny_tech · 2026-09-17
[source] Xiaomi livestreams MiMo-V2.6 RL training, burning over $1.08M in 36 hours — 量子位 · 2026-09-17
[source] Xiaomi opens MiMo-V2.6 live RL training panel: $1.16M+ spent, rewards still climbing — APPSO · 2026-09-17
Xiaomi live-streams MiMo RL training at 2B tokens per step; one user burns 2B tokens in 12 hours — MaziyarPanahi · 2026-09-17

2 near-duplicate retellings: burny_tech · burny_tech