Xiaomi's MiMo-V2.6-Pro Matches Grok 4.7 for $2.62 Million
A.I. / news
Xiaomi's MiMo-V2.6-Pro Matches Grok 4.7 for $2.62 Million
The MIT-licensed, trillion-parameter model tops Artificial Analysis' open-weight ranking and beats DeepSeek's V4.1-Flash on the same index, though Xiaomi's own numbers show it still trails Claude Opus 5 on some tasks.

The short version
Xiaomi released MiMo-V2.6-Pro and a smaller MiMo-V2.6-Flash on Sept. 21, open-sourcing both under the MIT license alongside a technical report disclosing what the reinforcement-learning run cost: about $2.62 million for Pro and $850,000 for Flash, across 30 RL steps and roughly 750,000 trajectories each, completed in under six days. Pro scored 46.32 on Artificial Analysis' Intelligence Index v4.3, tying newly released Grok 4.7 and topping the open-weight field, according to VentureBeat.
What 1.02 trillion parameters costs to train
MiMo-V2.6-Pro is a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token, a design that routes each token to a fraction of the network instead of running the whole thing. It supports a 1-million-token context window and takes text, image, video and audio in a single model, a combination Xiaomi calls "omnimodal." The company's disclosed RL cost breaks down as 43.5 percent on training, 43.8 percent on generating rollouts and 12.7 percent on grading them, according to VentureBeat's review of the technical report.
On Artificial Analysis' index, Pro's score of 46.32 sits ahead of DeepSeek's V4.1-Flash at 39 and DeepSeek's V4-Pro at 36, the model V4.1-Flash replaced in DeepSeek's own lineup last week. VentureBeat's Carl Franzen reported the figure directly from Artificial Analysis' published ranking rather than from Xiaomi's own marketing, which makes it one of the few numbers in this release that is not self-graded.
- MiMo-V2.6-Pro46.32 index points
- DeepSeek V4.1-Flash39 index points
- DeepSeek V4-Pro36 index points
Source: VentureBeat, accessed 2026-09-22
Flash is the model most people will actually run
MiMo-V2.6-Flash is the smaller of the two, at 309 billion total parameters and 15 billion active, priced at $0.14 per million input tokens and $0.28 per million output tokens against Pro's $0.435 and $0.87. Both are already live in Xiaomi's AI Studio, its MiMo apps, its own API and on OpenRouter.
| Model | Total / active parameters | Price per million tokens (in / out) |
|---|---|---|
| MiMo-V2.6-Pro | 1.02T / 42B | $0.435 / $0.87 |
| MiMo-V2.6-Flash | 309B / 15B | $0.14 / $0.28 |
Xiaomi is also releasing more than 7,000 reinforcement-learning task environments, the end-to-end RL framework it used to train both models, and MiMo-V2.6-Distill-Qwen-9B, a 9-billion-parameter model distilled from the larger models' RL trajectories onto a Qwen base, an unusual move for a company competing directly with Alibaba's Qwen line on the same leaderboard.
Where Xiaomi's numbers stop
VentureBeat's own writeup is explicit that "Xiaomi is not showing an across-the-board victory over the strongest closed models": Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol still lead on specific evaluations Xiaomi's own model card reports, including several agentic and coding benchmarks it lists as vendor-run rather than independently reproduced. Xiaomi's release material also promotes a feature it calls "Vibe World," letting the model build interactive 3D scenes from a text prompt and issue commands to a robotic arm, a capability neither VentureBeat nor Xiaomi's own benchmark tables attach a measured success rate to.
That gap, between a headline index score and the specific tasks a buyer would actually run, is the same one that trailed Alibaba's Qwen3.8-27B release in August, when an outside reviewer flagged Alibaba's self-reported comparisons as "select self-reported" rather than an apples-to-apples result. Xiaomi has not said when, or whether, an independent lab will run MiMo-V2.6-Pro through the same evaluations.
Sources
More in A.I.
- 01Qwen3.8-27B Hits 6.3 Million Downloads on Alibaba's Own Benchmark ClaimsThe weights are Apache 2.0 and ungated, but the card says nothing about training data, languages or the GPU needed to run them.
- 02Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are InImperial's Kevin Buzzard and MIT's Andrew Sutherland have examined the formalization, according to Scientific American. No journal review has been reported.
- 03Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 04Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.