Xiaomi MiMo published the MiMo-V2.6 series on September 21, 2026 — not a single checkpoint, but a same-day stack: open-weight Pro-RL and Flash-RL MoE models on Hugging Face, a separate 9B Qwen distill for agentic RL research, plus API SKUs on mimo.mi.com that include a latency-priced Pro-UltraSpeed tier.
According to Xiaomi's model cards and the accompanying technical report, both Pro and Flash are native omnimodal (text, image, video, audio) with a 1M-token context, trained with one mixed reinforcement-learning run across coding, general agents, visual work, and cybersecurity — a recipe the team brands as scaling RL toward self-improvement after weeks of public RL-dashboard livestreams.
Confirmed
- Three Hugging Face releases in one collection (MiMo-V2.6): MiMo-V2.6-Pro-RL (flagship MoE, MIT), MiMo-V2.6-Flash-RL (efficiency MoE, MIT), and MiMo-V2.6-Distill-Qwen-9B (SFT of Qwen3.5-9B on MiMo-generated data — a research starter, not a shrink of the Pro MoE).
- Architecture (vendor cards / report): Pro is a sparse MoE at 1.02T total / 42B activated; Flash is ~309–310B total / 15B activated. Both use hybrid sliding-window + global attention, a 681M MiMo ViT, audio encoders, and a 5-layer multi-token-prediction speculative decoder. Hugging Face "model size" labels (524B / 159B) are download-side parameter tags — use the total/active MoE figures for capacity comparisons.
- API product matrix (mimo.mi.com): V2.6-Pro list prices start at $0.0036 / $0.435 / $0.87 per MTok (cache hit / miss / output); Flash at $0.0028 / $0.14 / $0.28; Pro-UltraSpeed at $0.036 / $4.35 / $8.7 with a claimed up-to-20× inference-speed tier. Also marketed: MiMo Desktop, Batch API, Team Plan, and MiMo Claw (flagship Pro-powered office SKU).
- Vendor evaluation table (same numbers on Pro and Flash cards): on DeepSWE v1.1, Pro 71.9 / Flash 67.9 vs Claude Opus 5 74.0, GPT-5.6 Sol 73.0, Claude Fable 5 70.0. AutomationBench 53.1 / 52.3 edges the frontier rows Xiaomi lists. Cyber is split — CyberGym and MiMo Cyber Bench look very strong, while ExploitGym / ExploitBench / SEC Bench Pro trail Claude/GPT rows in the same table.
- Training story: asynchronous GRPO at large batch (1,568 prompts × 16 rollouts per step; billions of tokens per update), groupwise agentic grading (GRS/GAR), router freezing, and post-RL MOPD2 distillation. The report says Xiaomi is open-sourcing training dynamics, RL environments, and an RL framework alongside the Distill-9B checkpoint.
| Release | What it is | Scale (vendor) | Where |
|---|---|---|---|
| MiMo-V2.6-Pro-RL | Flagship omnimodal MoE (MIT) | 1.02T / 42B active · 1M ctx | Hugging Face / ModelScope · API |
| MiMo-V2.6-Flash-RL | Efficiency omnimodal MoE (MIT) | ~309B / 15B active · 1M ctx | Hugging Face / ModelScope · API |
| MiMo-V2.6-Distill-Qwen-9B | Qwen3.5-9B SFT on MiMo data (agent RL starter) | 9B dense | Hugging Face |
| MiMo-V2.6-Pro-UltraSpeed | API latency SKU (Pro-class quality claim) | Serving tier, not a separate HF MoE card | mimo.mi.com API |
Sources: Xiaomi MiMo HF collection, Pro/Flash model cards, mimo.mi.com product matrix — Sep 21, 2026.
| Benchmark | Pro | Flash | Opus 5 | GPT-5.6 Sol | Fable 5 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 74.0 | 73.0 | 70.0 |
| AutomationBench v1.0.6 | 53.1 | 52.3 | 50.3 | 45.8 | 46.2 |
| Toolathlon-Verified | 76.9 | 73.6 | 80.6 | 74.9 | 77.9 |
| Terminal Bench 4.0 | 34.9 | 28.8 | 49.0 | 39.9 | 42.4 |
| CyberGym | 94.0 | 95.1 | — | — | — |
| ExploitBench | 47.9 | 25.3 | 70.0 | 78.5 | 78.0 |
| MiMo VisualCoding | 72.3 | 71.5 | 70.0 | 73.4 | 69.1 |
Vendor table from XiaomiMiMo HF cards. Em dashes = not reported in that row. Several suites (MiMo Code/Cyber/Visual) are Xiaomi-internal.
Unknown
- Independent replication of DeepSWE, AutomationBench, CyberGym, and internal MiMo benches outside Xiaomi's harnesses.
- Whether Pro-UltraSpeed is a distinct weight file, speculative-decoding config, or dedicated serving cluster — only the 20× speed claim and ~10× list-price multiplier are published.
- Full open-source dump schedule for the RL environments/framework beyond the Distill-9B SFT starter and the cards' cookbook pointers (SGLang / vLLM).
- Hardware footprint and realistic tokens-per-dollar for self-hosting Pro (1.02T MoE) versus Flash at production concurrency.
- How the public RL livestream cost/step metrics map to the final released checkpoints.
Our take
Treat V2.6 as a portfolio drop: Pro/Flash are the real MoE story; Distill-9B is a research on-ramp on Qwen weights; UltraSpeed is a priced serving SKU. The interesting Brocker angle is the mixed-domain single RL run plus public training transparency — not every score in the launch table. Cyber especially: high CyberGym / MiMo Cyber without matching ExploitBench / SEC Bench Pro against Claude/GPT is a reminder that vendor cyber suites are not interchangeable. Route by task checkability (as with other China open-weight drops) and retest on your harness before trusting the DeepSWE near-frontier claim. Price/capability context: Chinese open models vs frontier API price–capability gap.
Sources
- Xiaomi MiMo: MiMo-V2.6 product page
- Xiaomi MiMo home — V2.6 series & API pricing
- Hugging Face: MiMo-V2.6 collection
- Hugging Face: MiMo-V2.6-Pro-RL
- Hugging Face: MiMo-V2.6-Flash-RL
- Hugging Face: MiMo-V2.6-Distill-Qwen-9B
- Xiaomi LLM-Core: MiMo-V2.6 technical report (PDF) — Scaling Reinforcement Learning Towards Self-Improvement, 44 pp., Sep 2026 (hosted on the Pro-RL HF repo).
