Xiaomi MiMo-V2.6-Pro Ties Grok 4.7 on the Intelligence Index at 46, Costs 45x Less Than Claude Opus 5, and Ships Under MIT

Xiaomi's MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index on Sep 22 - tied with SpaceXAI's same-day Grok 4.7, the highest open-weight result ever recorded - and ships under MIT (not research-only) at $0.435/M input. Lead developer is ex-DeepSeek's Luo Fuli. Same day Alibaba unveiled the Zhenwu V900.

September 22, 2026 - Xiaomi’s MiMo team dropped MiMo-V2.6-Pro today and it landed at 46 on the Artificial Analysis Intelligence Index v4.3 - the highest open-weight score ever recorded on the index, tied with SpaceXAI’s same-day Grok 4.7 release at 46, ahead of Grok 4.6 (44), Gemini 3.8 Flash (41), DeepSeek V4.1 Flash (39), and DeepSeek V4.1 Pro (36). The model is a 1.02T-total / 42B-active sparse mixture-of-experts with a 1M-token context window, ships open-weight under MIT (not the Qwen Research License Agreement that Qwen-Image-2.1 shipped under two days ago), and runs at $0.435 per million input tokens. The same-day news: Alibaba unveiled the Zhenwu V900 accelerator - China’s most powerful AI chip - on the same Tuesday. US export controls on Nvidia hardware were supposed to slow Chinese AI; a consumer-electronics company whose flagship product is a smartphone now ships the open-weight ceiling.

The Intelligence Index scoreboard: 46, open-weight ceiling, tied with Grok 4.7

The Artificial Analysis Intelligence Index v4.3 aggregates ten benchmarks into a single score. Treat it as the third-party aggregator the open-weight community has actually settled on - it has been measuring every major model release since Q4 2025 and its published numbers have tracked against the LMArena and Stanford HELM results closely enough that the gaps matter. The Sep 22 board, from Artificial Analysis’s same-day release:

Model AA Index v4.3 Weights Notes
MiMo-V2.6-Pro 46 open (MIT) Xiaomi, 2026-09-22
Grok 4.7 46 closed SpaceXAI, 2026-09-22
Grok 4.6 44 closed prior SpaceXAI release
Gemini 3.8 Flash 41 closed Google, Sep 2026
DeepSeek V4.1 Flash 39 open (Apache 2.0) DeepSeek, 2026-09
DeepSeek V4.1 Pro 36 open (Apache 2.0) DeepSeek, 2026-09
Claude Opus 5 ~44 (5 points behind Pro on per-task economics) closed Anthropic, 2026
GPT-5.6 Sol 43 closed OpenAI, Aug 2026

The story is two lines. First: Xiaomi, a Chinese consumer-electronics company better known for smartphones and EVs, ships the open-weight ceiling at the same number SpaceXAI ships closed weights. Second: that ceiling is now 11 points above the strongest Apache-2.0 Chinese competitor (DeepSeek V4.1 Pro at 36) and only 5 points behind Anthropic Opus 5 on the per-task economics Artificial Analysis measures. The Intelligence Index moved more in the last 30 days than in the prior six months.

The architecture: 1.02T total / 42B active, 1M context, hybrid attention with MTP

The MoE numbers are not new - Xiaomi’s MiMo-V2.5-Pro was already 1.02T total / 42B active in the prior generation. MiMo-V2.6-Pro is an iteration, not a parameter-count jump. What changed:

  • 1.02T total / 42B active sparse MoE - confirmed across Xiaomi’s official docs, Wikipedia, and Kingy.ai’s spec table
  • 1M-token context window with 131K max output
  • Native multimodal input - text, image, audio, video in a single forward pass; not a bolted-on vision adapter
  • Hybrid attention with a 5-layer MTP (Multi-Token Prediction) speculative decoder for inference-time acceleration
  • MiMo-V2.6-Pro-UltraSpeed - a separately deployed variant that generates at up to 20x the standard Pro output speed for high-throughput serving

The MTP decoder is the inference-economics lever. Speculative decoding with five parallel draft tokens means a single H200 (or, increasingly, a Chinese accelerator like the V900 being unveiled today) can serve more concurrent sessions per chip. The 1M context is also structural - prior 1M-context models (Magic 2.0, Ling-2.5 Ring) achieved that ceiling with quadratic-attention compromises; MiMo-V2.6-Pro is the first open-weight 1M-context model with hybrid (linear + sliding-window) attention published in the official docs.

The license: MIT, not Qwen Research, not Llama Community

This is the structural line item. From Xiaomi’s official Hugging Face collection page and VentureBeat’s launch report:

The model weights are released under the MIT License. The license applies to model weights, code, and documentation.

MIT, in full. That means:

  • Commercial use - ship it in a SaaS product, embed it in a paid agent, resell inference - all permitted
  • Modification - fine-tune, merge, distill, quantize - permitted
  • Redistribution - host your own fork, redistribute derivatives - permitted, with copyright notice preserved
  • No use-based restrictions - no “non-commercial,” no “research only,” no “below 700M MAU” ceiling

Compare to the open-weight releases from the same Chinese ecosystem in the prior 30 days:

Release Date License Commercial use
DeepSeek V4.1 Pro 2026-09-09 Apache 2.0 Yes
Qwen-Image-2.1 2026-09-20 Qwen Research License Agreement No (separate written agreement required)
Moonshot Muse Spark 1.3 2026-09-15 Modified MIT with attribution Yes
Xiaomi MiMo-V2.6-Pro 2026-09-22 MIT (full) Yes
Llama 5 Anvil 2026-08-30 Llama Community License Yes (with 700M MAU clause)

The Chinese open-weight ecosystem is not a monolith on licensing. Alibaba moved Qwen-Image-2.1 to research-only two days ago; Xiaomi shipped MiMo-V2.6-Pro under full MIT the same week. DeepSeek stays Apache 2.0. For an indie builder or SaaS founder deciding which open-weight model to standardize on, license is now a primary variable, not a footnote - and the “MIT” license is the broadest of the three.

The pricing: $0.435/M input, $0.13 per Intelligence Index task, 45x cheaper than Opus 5

Three prices matter for the agent-stack economics:

Tier Input Output Notes
MiMo-V2.6-Pro $0.435 / M tokens $0.87 / M tokens API, uncached
MiMo-V2.6-Flash $0.14 / M tokens $0.28 / M tokens smaller, same family
MiMo-V2.6-Pro-UltraSpeed $0.435 / M (input) higher output rate, throughput-optimized up to 20x output speed

The figure that lands: $0.13 per Artificial Analysis Intelligence Index task, measured by Artificial Analysis. Claude Opus 5 measured at $5.86 per Intelligence Index task on the same protocol. That is a 45x price gap on identical task difficulty. The score gap between Pro and Opus 5 on the same protocol is 5 points on the index, with Pro winning on Terminal-Bench 2.1 (89.9% vs Opus 5 lower), CyberGym (94.0%, leads all reported models), and AutomationBench v1.0.6 (53.1%), narrowly trailing on DeepSWE v1.1 (71.9% vs 74.0%) and a handful of others.

For the hands-on economic comparison: Kingy.ai ran three coding/data tasks with 23 hidden automated checks against both MiMo-V2.6-Pro and Claude Opus 5 through OpenCode. Both passed 23/23. MiMo-Pro cost $0.03 for the run. Opus 5 cost $1.03. Same accuracy, 34x cheaper.

The benchmark table vs Opus 5: 4 wins, 1 narrow loss, the fair read

Xiaomi publishes its own benchmark comparison. Treat them as vendor-published; the hands-on Kingy.ai data is the independent confirmation. From Forkast News’s same-day report and Xiaomi’s official model card:

Benchmark MiMo-V2.6-Pro Claude Opus 5 Verdict
Terminal-Bench 2.1 89.9% lower Pro wins
CyberGym 94.0% lower Pro wins (leads all reported models)
AutomationBench v1.0.6 53.1% lower Pro wins
SWE-Bench Verified (Thinking) 78.6 higher Opus 5 wins
DeepSWE v1.1 71.9% 74.0% Opus 5 wins (narrow)

Xiaomi’s own comparison table has MiMo-V2.6-Pro trailing Opus 5 on 10 of 14 shared evaluations. The fair read: close-to-frontier at a fraction of the cost, not across-the-board parity. The agent-stack takeaway is the price-per-task ratio, not a blanket “MiMo beats Opus 5.” For the developer building a coding agent who needs Terminal-Bench / CyberGym / AutomationBench class capability at sub-cent cost, MiMo-V2.6-Pro is the strongest open-weight option shipping today.

The same-day chip: Alibaba Zhenwu V900, Luo Fuli, and the export-control counter-narrative

The Tuesday timing is not an accident. The same day Xiaomi shipped MiMo-V2.6-Pro, Alibaba unveiled the Zhenwu V900 accelerator - described in Alibaba’s press materials as China’s most powerful AI chip, with 500K-chip cluster buildout planned for Q1 2027, a 10-trillion-parameter model roadmap, and 20GW cloud datacenter capacity by 2032. The AP wire, Time News, and Forkast News all covered both drops on Sep 22.

The structural shift: US export controls on Nvidia hardware (H100/A100 restrictions from 2022, H200 blocked January 2026, third-country cloud loopholes closed May 2026) were intended to slow Chinese AI development. The empirical record is the opposite. From Forkast News’s same-day analysis:

  • Chinese accelerator market share: near-zero in 2022 → 41% in 2025 (domestic chips like Huawei Ascend, Cambricon, T-Head, and now Zhenwu V900)
  • Nvidia China datacenter revenue: effectively zero in Q2 2026 vs the multi-billion-dollar baseline of 2023-2024
  • A consumer-electronics company now ships the open-weight ceiling the same day a Chinese chip vendor unveils a domestic accelerator that can train one

Lead developer Luo Fuli (formerly DeepSeek, joined Xiaomi late 2025) said in a KuCoin News interview that the development challenges for MiMo-V2.6 surpassed those of DeepSeek R1, and committed to open-sourcing the training stack and RL environments in pieces over the coming weeks. The training stack release is the second-order story: if Xiaomi ships the RL environments and reward-model code under MIT, the Chinese open-weight ecosystem will have its first full-stack open recipe for frontier-class training, distinct from DeepSeek’s earlier (still-partial) releases.

What it means for builders, buyers, and tinkerers

Three concrete takeaways:

  1. If you are an indie builder or agent-stack developer: pull XiaomiMiMo/MiMo-V2.6-Pro-RL from Hugging Face today. MIT license means you can fine-tune, deploy, and ship it in a paid product without contacting Xiaomi. The 1.02T/42B MoE will not fit on a single consumer GPU, but 4-bit quantizations at 42B-active are landing within 48 hours on Hugging Face based on the community response pattern from prior DeepSeek and Qwen releases. PrismML Bonsai 2 27B shipped ternary (1.58-bit) at 27B-active and ran on a 4090 - the same community will quant MiMo-V2.6-Pro within a week.

  2. If you are a SaaS founder choosing between MiMo-Pro and Opus 5: MiMo-Pro wins on price-per-task by 34x to 45x on the same accuracy in independent hands-on testing, wins on three published benchmarks, ties on SWE-Bench class work within a few points, and ships under MIT so you do not need to negotiate a commercial agreement. The case for Opus 5 is the residual edge on the 10 of 14 benchmarks Xiaomi publishes where it still leads. For a coding-agent or data-pipeline product where Terminal-Bench / CyberGym / AutomationBench are the right targets, MiMo-Pro is the economic answer.

  3. If you are tracking the open-weight ecosystem: the Chinese stack just produced its strongest single release of 2026 - MIT-licensed, frontier-tied, same-day chip drop, open training stack on the way. The September closed-lab cohort that the Sep 16 post covered (Anthropic Mythos 5.1, OpenAI GPT-6 Astra, Gemini 3.8 Flash, Muse Spark 1.3, DeepSeek V4.1-Flash) just got out-benched on the open-weight side. The structural read: the export-control strategy accelerated, not slowed, the Chinese AI stack, and a consumer-electronics company is now shipping the highest-scoring open weights on the third-party aggregator the community uses.

Sources