XiaomiMiMo/MiMo-V2.6-Pro-RL

TL;DR: Xiaomi's MiMo-V2.6-Pro-RL is a 1.02T-parameter sparse MoE (42B active) with native text/image/video/audio input and 1M-token context, released under MIT. The interesting part isn't the scale — it's the RL recipe: one mixed RL run across coding, agents, visual, and cybersecurity domains, plus a "groupwise agentic grading" scheme where the model's own rollouts are compared against each other to synthesize rewards, closing a self-improvement loop. Reported numbers are within a few points of the top proprietary systems on agent benchmarks, but everything is vendor-reported and self-hosting requires a multi-node cluster.

What it is

The flagship of the MiMo-V2.6 series. Architecture: 70 transformer layers (60 sliding-window attention, 10 global), 384 routed experts with 8 active, no shared experts, hidden size 6144. A 681M-param ViT handles image/video (2×16×16 patches, mostly SWA layers), and audio comes through a 308M RVQ tokenizer plus a 127M patch encoder that downsamples 25 Hz → 6.25 Hz. A 5-layer MTP drafter predicts 7 tokens per forward pass for speculative decoding. There's also a Flash variant (~309B per the selection notes) that's the realistic self-host option.

The RL recipe is the actual contribution

Three ideas stand out from the card:

"You Only RL Once" — instead of separate per-domain RL runs, coding, general-agent, visual, and cybersecurity tasks (across multiple harnesses) are mixed in the same GRPO batch. The claim is cross-domain transfer, including to harnesses never seen in training. Batches are enormous: 1,568 prompts × 16 rollouts per step, fully async.

Groupwise agentic grading — the key framing: binary pass/fail can't rank passing solutions, so they scale the reward signal itself. Offline, Groupwise Reward Synthesis (GRS) builds task-specific rubrics by contrasting rollouts within a group, fused with test outcomes. Online, Groupwise Advantage Redistribution (GAR) ranks passing trajectories and shifts advantage toward higher-quality (shorter, cheaper) solutions. Because the grader judges the policy against its own samples, the reward model improves as the policy does — that's the "self-improvement" claim.

MOPD2 — after mixed RL, multi-prefix multi-teacher on-policy distillation: student rollouts conditioned on prefixes taken from teacher trajectories and SFT demos, so decision points get trained without regenerating whole multi-turn histories. Aimed at hard-to-verify tasks that pure RL can't reward.

Reported results

Self-reported, against proprietary frontier models. The jump from V2.5 is dramatic on agent tasks:

Pattern: parity or near-parity with the best closed models on most agent benchmarks, ahead on AutomationBench and MiMo VisualCoding, notably behind on Terminal Bench 4.0 (34.9 vs 49.0) and exploit-development benchmarks. The 3–4× gains over V2.5 on identical benchmarks are the strongest internal evidence the RL recipe does something real.

Caveats

• All benchmarks are self-reported. No independent evals yet, and downloads show 0 at time of writing — nobody has verified anything.
• Impractical to self-host: the reference SGLang config is tp16 × dp2 across 2 nodes — roughly 32 GPUs even at FP8. The Flash variant is the one to actually run.
• No training details beyond the card: the technical report link exists but the recipe (GRS rubric construction, GAR ranking, MOPD2) is described at the paragraph level. Reproducibility is unclear.
• Omnimodal input; nothing suggests audio/image output.
• MIT license, which is generous at this scale.

Try it

Realistically, use the hosted API (MiMo Studio, OpenRouter) unless you have a cluster. For self-hosting the smaller footprint:

docker pull vllm/vllm-openai:mimov25-cu129
vllm serve XiaomiMiMo/MiMo-V2.6-Pro-RL \
  --tensor-parallel-size 8 --trust-remote-code \
  --gpu-memory-utilization 0.95 --max-model-len auto \
  --reasoning-parser mimo --tool-call-parser mimo \
  --enable-auto-tool-choice
# sampling: temperature=1.0, top_p=0.95

Model card: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL — and check the sibling MiMo-V2.6-Flash-RL repo first if you want to run it yourself.