Xiaomi MiMo-V2.6: Livestreaming a Trillion-Parameter RL Training Run

September 17, 2026 — Fuli Luo, who leads Xiaomi's MiMo AI team (and previously worked at DeepSeek), announced that MiMo-V2.6 is in the middle of a large-scale reinforcement learning (RL) training run — and that the team is livestreaming it publicly. The announcement follows what Luo described as "nearly half a year of silence," spent studying one specific question: how far can RL actually scale?

TL;DR — what people are asking

Why "nearly half a year of silence" is the actual news

Luo's framing is unusual for a model announcement — most launches lead with capability claims, not a description of what the team spent six months not shipping. "We spent it studying one problem: how far RL can scale," she wrote, positioning MiMo-V2.6 less as a product launch and more as the output of a sustained research bet on reinforcement learning as the primary lever for further model improvement, rather than continuing to scale pretraining compute alone.

That framing matters because it's an implicit claim about where the field's marginal gains are coming from in late 2026: not bigger pretraining runs, but bigger, better-designed RL runs on top of an existing base model. It's the same directional bet explainx.ai has tracked across Xiaomi's earlier robotics foundation-model work and the broader industry shift toward agentic RL as the dominant post-training paradigm.

The three things MiMo scaled

Luo's thread names three specific scaling axes, each with a concrete number attached — a level of specificity that's genuinely useful for anyone trying to replicate or benchmark against this work:

The "environments and harnesses" axis is worth dwelling on. Mixing multiple agentic harnesses inside one RL run — rather than training separately per-harness and merging afterward — implies the team is trying to build a single policy that generalizes across different tool-use and task structures simultaneously, rather than a narrower specialist model. That's a materially harder training-engineering problem than single-harness RL, since reward signals, episode lengths, and action spaces can differ substantially between harnesses.

The technical pushback worth taking seriously

Not every reply was celebratory. Several raised specific, substantive engineering questions rather than general skepticism:

• Yifan Jiang challenged the "fully async" claim directly: "trainer time + sampler time = step time, so it's not 'fully async'" — pointing out that if the reported step time is simply the sum of trainer and sampler time, that's evidence of at least some sequential dependency between the two phases, not true overlap. The MiMo team hasn't yet published a response clarifying the actual pipeline overlap.
• Dinghuai Zhang questioned the design choice to run "flash RL" and "pro RL" simultaneously, without an explained rationale in the thread itself.
• One independent developer, replying as LLMWildling, reported doing comparable work independently at smaller scale: "122B running 16 rollouts at 200k context" — a useful external data point suggesting the general RL-scaling recipe (large rollout counts, long context) is being explored by more than one team concurrently, at different budget tiers.

None of these amount to a debunking — they're the kind of detailed technical questions a livestreamed, in-progress run invites precisely because it's unusually transparent. That transparency is itself the interesting part: publishing training internals live, rather than only in a post-hoc paper, opens the process to exactly this kind of real-time scrutiny.

Open-sourcing "piece by piece"

Luo committed to releasing the methodology incrementally — "we'll open-source the details piece by piece over the coming weeks" — rather than as a single comprehensive technical report. That's a different release cadence than the typical single-paper-plus-weights drop from labs like Moonshot AI or Alibaba's Qwen team, and it means anyone tracking this story should expect follow-up posts rather than one definitive writeup.

This continues a pattern of unusually open compute disclosure from Chinese AI labs in 2026 that explainx.ai has tracked closely — see the detailed parameter and training disclosures around Kimi K3's 2.8 trillion parameter open-weights release and Qwen 3.8 Max's open-weight token strategy — a contrast worth noting against the comparatively guarded training-compute disclosures typical of OpenAI, Anthropic, and Google DeepMind.

Why "grader compute" is its own scaling axis, not an afterthought

Most public discussion of RL scaling focuses on the two more intuitive axes — how much compute goes into generating rollouts, and how diverse the training environments are. The MiMo team's explicit call-out of grader compute as a third, independently scaled axis is worth dwelling on, because it's frequently the least-discussed bottleneck in agentic RL despite arguably mattering as much as the other two. A rollout is only as useful for training as the reward signal attached to it — if the grading step (deciding how good a given rollout actually was) is cheap, noisy, or coarse, no amount of rollout-generation compute compensates for a weak or miscalibrated signal.

"Agentic in-group credit assignment" combined with "test-case and rubric-based rewards" suggests the MiMo team is running its own agentic evaluation process to score each rollout relative to others in its group — effectively spending real inference compute on the grading step itself, not just on generating candidate behavior. That's a materially more expensive approach than a simple binary pass/fail reward or a lightweight heuristic score, and treating it as a first-class scaling axis (alongside rollout compute and environment diversity) implies the team's internal experiments found grading quality to be a genuine bottleneck worth spending compute against directly, rather than an implementation detail to minimize cost on.

What "livestreaming a training run" actually reveals — and doesn't

It's worth being precise about what a public training stream can and can't show an outside observer. Live metrics like loss curves, reward trends, or rollout throughput give real-time visibility into whether training is proceeding stably — useful for anyone trying to diagnose their own RL runs against a working reference. What a stream can't show is the accumulated design decisions that got the team to a working configuration in the first place: which reward formulations were tried and discarded, which environment mixes failed to converge, or what infrastructure bugs consumed weeks before the currently-streaming run began. Luo's framing of "nearly half a year of silence" spent on this exact question suggests a substantial amount of that trial-and-error already happened before the public stream started — meaning what's being livestreamed is closer to a validated configuration in its execution phase than an from-scratch, watch-everything-including-the-failures research process. That's still valuable transparency, but it's a different kind of transparency than showing the messy middle of the research itself.

Honest limitations

• No benchmark results yet. This is a training run in progress, not a released model — MiMo-V2.6's actual capability level relative to competitors is unknown until training completes and evaluations are published.
• The "fully async" claim has an unresolved technical objection from the community that the MiMo team hasn't yet addressed publicly.
• "Piece by piece" open-sourcing has no committed timeline beyond "over the coming weeks" — treat this as an ongoing story, not a closed one.
• No parameter count for MiMo-V2.6 itself is confirmed in this specific thread, distinct from the separate community report of a 1-trillion-parameter MiMo RL run circulating the same day.
• Livestreaming a training run is a form of marketing as well as transparency — it generates sustained public attention across the training period, which is a real incentive independent of the scientific value of openness.

What this means for builders and researchers

Teams doing their own RL post-training: the specific numbers here — 1,568 prompts × 16 rollouts, ~2B tokens/step, mixed multi-harness environments — are a useful reference point for scaling your own agentic RL setup, even before MiMo-V2.6's final benchmarks land.

Anyone evaluating open vs. closed model strategy: this is another data point in the pattern of Chinese labs treating detailed training methodology disclosure as a competitive differentiator rather than a risk, worth weighing against closed-lab alternatives when picking a model family to build on.

Infrastructure engineers: the "fully async" debate in the replies is worth following even independent of MiMo-V2.6 specifically — trainer/sampler overlap is a genuine, generalizable RL-infrastructure problem, and a clear public answer from the MiMo team (if one comes) would be useful reference material regardless of which framework you use.

Related on explainx.ai

• Kimi K3 open weights: 2.8 trillion parameters
• Qwen 3.8 Max preview: open-weight token strategy
• Xiaomi Robotics U0: world foundation model via embodied synthesis
• Scalable oversight: RLHF, Constitutional AI, weak-to-strong generalization
• Top Chinese AI companies and startups guide
• Moonshot AI detention rumor and the Anthropic distillation theory

Official source: @_LuoFuli on X (September 17, 2026); live training stream at mimo.xiaomi.com/rl/

Details reflect Xiaomi MiMo's September 17, 2026 announcement of an in-progress training run — final model specifications and benchmark results were not yet published as of this writing.