MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

What it claims

Xiaomi's LLM-Core team reports two omni-modal (text, image, video, audio) Mixture-of-Experts models trained with heavily scaled RL:

• MiMo-V2.6-Pro: 1.02T total parameters, 42B active.
• MiMo-V2.6-Flash: 310B total parameters, 15B active.

The open-source release is narrower than the model series. It is a distilled MiMo-V2.6-Distill-Qwen-9B, curated task environments with verifiers across several domains, an end-to-end RL training framework, a composable "mini-harness" for agentic interaction, and the training dynamics. The extracted text does not say that Pro or Flash weights are released.

How it works

Backbone

• Hybrid blocks: N sliding-window attention (SWA) layers followed by one global-attention (GA) layer. The window is 128.
• The first block is GA with a dense FFN. All other FFNs are sparse MoE with no shared experts.
• Flash has 48 layers (39 SWA, 9 GA), hidden size 4096, and 256 experts with 8 active.
• Pro has 70 layers (60 SWA, 10 GA), hidden size 6144, and 384 experts with 8 active.

Encoders

• MiMo-ViT (681M parameters, 28 layers, 24 SWA and 4 GA) uses sink-augmented SWA. It alternates row-major and column-major token order. It was pre-trained from scratch on more than 4T image tokens, paired with a small LLM that stays trainable, using plain cross-entropy and no contrastive objective.
• Audio: an RVQ tokenizer (308M parameters) produces 20 codebooks at 25 Hz and was trained on 20M hours of audio. A patch encoder (127M parameters) groups 4 frames per patch, reducing the rate to 6.25 Hz for the backbone.

Speculative decoding

• A 5-layer, all-SWA drafter (window 1024) follows the DFlash block-diffusion design.
• It predicts 7 tokens per forward pass for parallel verification by the backbone.

Pre-training

• Two stages: text-only first, then joint omni-modal training. Context starts at 32K and is extended to 256K partway through. The optimizer is AdamW.
• Flash: 48T tokens (26T text, 22T omni).
• Pro: 30T tokens (27T text, only 3T omni).

Mid-training

• Agent-centric data mixture. Most compute runs at 256K context, with a final stage extending to 1M.
• The hidden weight matrices switch from AdamW to Muown, a Muon variant with row-norm control. The motivation is that AdamW's optimization efficiency diminished at large mixed-task RL batch sizes. Embeddings, the LM head and the MoE router stay on AdamW.
• Prior work warned that switching an Adam-pretrained model to Muon causes optimizer mismatch. The authors report no loss spike.
• MXFP4 quantization-aware training is applied during mid-training.

Groupwise grading

• Passing solutions are not all given the same reward. Instead, solutions within a group are compared against each other.
• These comparisons are turned into learning signals through Groupwise Reward Synthesis (GRS) and Groupwise Advantage Redistribution (GAR).

RL infrastructure

• A unified trajectory representation with a penalty mechanism.
• A harness pool for high-concurrency rollout across multiple agent frameworks.
• Control and data planes decoupled, so large trajectories carrying routing and multimodal payloads can be buffered and transferred.
• A sample mixer combined with dynamic sampling and partial rollout, to keep per-task batch composition stable.
• Training and inference engines aligned on MoE routing and on top-p candidate sets.

Evidence

• Figure 1 plots scores over roughly 30 RL steps for both models on six benchmarks: DeepSWE v1.1, SWE-Bench Pro, MiMo Code Bench, AutomationBench v1.0.6, MiMo Visual Coding and MiMo Cyber Bench. Scores rise monotonically.
• The axis ranges are visible, e.g. DeepSWE roughly 48→72 and MiMo Cyber Bench roughly 66→78. The extracted text gives no per-model final numbers and no comparison against external models.
• Three of the six benchmarks are in-house: MiMo Code Bench, MiMo Visual Coding and MiMo Cyber Bench.
• The authors say RL experiments on the 9B distilled model show consistent gains across tasks and harnesses. No figures appear in the extracted text.

How to use it

The realistic entry point for most engineers is the open suite rather than the frontier models:

• MiMo-V2.6-Distill-Qwen-9B.
• The verifier-backed environments.
• The RL framework.
• The mini-harness.

The page gives no repository URLs, install commands or licence. Check those before planning around the release.

Limits

• The fetched text stops at the start of Section 4. Everything after that point is visible only as section headings. This includes:
• Details of GRS and GAR.
• The reward-hacking defenses.
• The router-freezing ablation (§5.4).
• RL failure analysis (§5.5).
• MOPD² distillation.
• All quantitative results tables.
• Benchmark claims rely partly on Xiaomi-internal benchmarks.
• Pro received far less omni-modal pre-training than Flash (3T vs 22T tokens). That may matter for multimodal use of Pro, but the page does not evaluate this.