Xiaomi’s MiMo-V2.6 launch post (update time September 22, 2026) frames the release as scaling reinforcement learning for self-improvement — and as a full open-source drop: weights, technical report, and RL supporting resources on the Hugging Face MiMo-V2.6 collection.
What Xiaomi published for MiMo-V2.6
Per Xiaomi’s primary post, the series comprises two native fully multimodal models — Pro and Flash — covering text, image, video, and audio in one stack. Xiaomi states MiMo-V2.6-Pro scores 46 points on the Artificial Analysis Intelligence Index (AA Composite Intelligence Index), surpassing Kimi K3 and Qwen3.8 Max to become, in their framing, the most powerful open-source model available — while still leaving a gap versus closed-source Claude Fable 5.1 and GPT-6 Astra. Those are vendor-attributed claims from the Xiaomi post, not this desk’s independent run.
The company also released MiMo-V2.6-Distill-Qwen-9B and open-sourced RL task environments, an end-to-end RL training framework, and lightweight harnesses. Live RL training notes on the post cite multi-day runs and trajectory counts; treat those as Xiaomi’s own training diary, not Alberta capacity planning.
“Benefiting from the expanded RL computing power, MiMo-V2.6-Pro scores 46 points in the Artificial Analysis Intelligence Index (AA Composite Intelligence Index), surpassing Kimi K3 and Qwen3.8 Max to become the most powerful open-source model available; However, there is still a gap when compared with the strongest closed-source models Claude Fable 5.1 and GPT-6 Astra.”Xiaomi MiMo — MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement, September 22, 2026
Params, license, and API pricing (verified)
Hugging Face model cards under the MiMo-V2.6 collection list license: mit for both MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL. Model Summaries on those cards state sparse MoE architecture with Pro: 1.02T total / 42B activated and Flash: 309B total / 15B activated, both with 1M token context and text/image/video/audio modalities. The collection listing also shows “model size” chips of 524B (Pro) and 159B (Flash); this desk treats the Model Summary total/activated figures as the authoritative architecture claim and notes the listing chips as Hugging Face’s separate size display.
Xiaomi states the MiMo-V2.6 series adopts the same API pricing as the V2.5 series — intelligence up, list price unchanged. The news page renders pricing as an image, not copy-pasteable text. This briefing therefore does not invent dollar amounts. Confirm live rates on Xiaomi’s platform token plan before budgeting; quote only “same list pricing as MiMo-V2.5 / unchanged per Xiaomi” until you read the rate card yourself. API callers must use lowercase names: mimo-v2.6-pro, mimo-v2.6-flash, and mimo-v2.6-pro-ultraspeed (UltraSpeed described as up to 20× inference speed on Xiaomi’s post).
| Line | Claim | Source | Notes |
|---|---|---|---|
| Pro architecture | 1.02T / 42B active | HF Model Summary | Sparse MoE · 1M ctx · MIT |
| Flash architecture | 309B / 15B active | HF Model Summary | Sparse MoE · 1M ctx · MIT |
| AA Composite (vendor) | Pro = 46 | Xiaomi news | Ahead of Kimi K3 / Qwen3.8 Max per Xiaomi; behind Fable 5.1 / GPT-6 Astra |
| API list pricing | Same as V2.5 | Xiaomi news | Do not invent $; check platform rate card |
What Alberta operators should do
For private AI security and industrial AI Alberta shops, MIT open weights matter more than a hosted demo. If the workload is tenders, plant telemetry, client credentials, or shop-floor SOP agents, a foreign hosted API is not a private Alberta path — even when the list price looks attractive. Flash’s 15B activated MoE is the more realistic self-host candidate to size first; Pro’s 1.02T total is a serious multi-GPU (or research-cluster) bill before you celebrate the leaderboard. Use the API names above for a short hosted eval, then decide whether weights leave Xiaomi’s cloud.
Pair this briefing with Opcelerate’s private AI security lane and AI consulting Alberta path when the stack touches regulated or customer data. Keep reading the city desk at The Super Intelligence Times for source-backed model cards — not screenshots with invented rates. Hosted Xiaomi API ≠ on-prem Alberta inference. Put a human review gate on any agent that can write tickets, change MES setpoints, or email customers.
The guardrail
Open weights are not a free pass. MIT lets you run, modify, and commercialize — it does not verify Xiaomi’s agent benchmarks on your harness, and it does not place GPUs in Sherwood Park. Do not treat “most powerful open-source” marketing as proof of AGI or ASI. Size VRAM from the HF cards, confirm the rate card if you stay on API, and keep sensitive industrial workflows on a private agents path with logging and ownership.