MiMo-V2.6-Pro-RL

What it claims

MiMo-V2.6-Pro-RL is the flagship checkpoint of Xiaomi's MiMo-V2.6 series. Its training thesis is that RL compute, environment diversity and grader compute should all scale together. GRPO runs fully asynchronously at 1,568 prompts × 16 rollouts per step, which the card describes as "billions of tokens per update". Tasks and multiple agent harnesses share each batch. The card claims this mixing lets learned strategies transfer to harnesses never seen in training, and that the groupwise grader pushes the model toward shorter paths and fewer tokens per task. It gives no numbers for either claim.

How it works

Training has three stages beyond pretraining:

flowchart LR
  CS[Cold start: self-correction] --> RL[Mixed async GRPO]
  RL --> R[Group of rollouts]
  R --> GRS[GRS: offline rubrics + test outcomes]
  R --> GAR[GAR: rank passing rollouts]
  GRS --> ADV[Advantage]
  GAR --> ADV
  ADV --> RL
  RL --> M[MOPD2]
  S[Student rollouts] --> M
  TP[Teacher-Prefix] --> M
  SP[SFT-Prefix] --> M

• Cold start. The model rewrites its own misaligned turns into grounded next steps. During RL, environment hardening, adversarial screening and verifier cross-checks are used against reward hacking.
• Groupwise grading. Binary pass/fail cannot separate passing solutions, so an agentic grader compares rollouts within each group:
• GRS builds task-specific rubrics offline from contrasting rollouts, then fuses rubric quality with test results.
• GAR shifts advantage online toward the better passing trajectories.
• MOPD2. Student rollouts are mixed with single-turn rollouts that are conditioned on a prefix taken from teacher trajectories or SFT demonstrations. This trains individual decision points without regenerating the earlier turns. It is aimed at tasks that are hard to verify.

Architecture details not in the summary:

• Backbone. Block 1 uses global attention with a dense FFN. All later blocks interleave sliding-window and global attention, with MoE FFNs that have no shared experts.
• Vision. A 681M-parameter ViT with 28 layers: 24 sliding-window and 4 full attention.
• Audio, tokenizer. A 308M-parameter AudioTokenizer: 24 layers (12 sliding-window, 12 global), hidden size 1024, 20 RVQ codebooks.
• Audio, patch encoder. 127M parameters and 6 layers, packing 4 frames per patch (25 Hz → 6.25 Hz).
• Speculative decoding. The MTP drafter is DFlash-style with 5 sliding-window layers and predicts 7 tokens per forward pass.

Evidence

The card lists evaluation entries for:

• Deep-SWE
• Terminal-Bench 2.1
• Toolathlon Verified
• "Terminalbench 4"

The page text shows no scores, baselines or settings, so there is nothing here to compare against.

How to use it

Serving needs multi-node scale. The reference SGLang setup spans 2 nodes with TP 16, DP 2, EP 16, DeepEP all-to-all and EAGLE speculative decoding. The card points to the "SGLang MiMo cookbook" for best performance and uses Docker image lmsysorg/sglang:latest.

sglang serve \
  --trust-remote-code \
  --model-path XiaomiMiMo/MiMo-V2.6-Pro-RL \
  --tp 16 \
  --dp 2 \
  --enable-dp-attention \
  --mm-enable-dp-encoder \
  --ep 16 \
  --moe-a2a-backend deepep \
  --moe-dense-tp-size 1 \
  --mem-fraction-static 0.7 \
  --max-running-requests 128 \
  --chunked-prefill-size 32768 \
  --page-size 64 \
  --swa-full-tokens-ratio 0.3 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --enable-multi-layer-eagle \
  --reasoning-parser mimo \
  --tool-call-parser mimo \
  --host 0.0.0.0 \
  --port 30000 \
  --nnodes 2 \
  --node-rank <node-rank> \
  --dist-init-addr <node0-ip>:20000

For vLLM, the card points to the V2.5 recipe and image (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 \
  --generation-config vllm

• Sampling. Recommended settings are temperature=1.0, top_p=0.95.
• Remote code. trust_remote_code is required.
• Hosted access. The model is also offered through AI Studio, MiMo Code, Xiaomi MiMo Desktop, the MiMo Open Platform API and OpenRouter.

Limits

• No benchmark numbers. The fetched text has none, and it does not quantify the claims about cross-harness transfer or token efficiency.
• No licence stated. Check the repository files before adopting.
• No GPU spec. The card never names the GPU type or memory needed for either serving configuration.
• vLLM path is borrowed. It relies on a V2.5 recipe and image, so V2.6-specific support is unconfirmed.
• Two drafter descriptions. The SGLang config uses EAGLE with 3 steps and 4 draft tokens, while the drafter is described as predicting 7 tokens. The card does not say how these relate.