Xiaomi MiMo-V2.6 Tops Open-Weight Rankings With 1T-Param Model
Xiaomi MiMo-V2.6 Tops Open-Weight Rankings With 1T-Param Model
Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash on September 22, 2026, putting the top spot on the Artificial Analysis open-weight intelligence leaderboard under the MIT license for the first time. At $0.87 per million output tokens for Pro and just $0.28 for Flash, the models offer enterprise teams a credible alternative to proprietary frontier models at a fraction of the cost.
What Launched
Xiaomi’s official MiMo announcement confirms two new models in its MiMo series, both built as native omnimodal sparse mixture-of-experts architectures trained primarily through reinforcement learning:
- MiMo-V2.6-Pro: 1.02 trillion total parameters, 42 billion active per forward pass, 1-million-token context, text plus image plus video plus audio input, MIT license.
- MiMo-V2.6-Flash: 309 billion total parameters, 15 billion active, the same 1-million-token context and modality support, also MIT licensed.
Both ship under the MIT license. Weights are available on Hugging Face and hosted access is available through the Xiaomi API, OpenRouter, and Xiaomi AI Studio.
Xiaomi is also releasing MiMo-V2.6-Distill-Qwen-9B, a 9B-parameter model fine-tuned using data generated by the V2.6 series, and a third API-only tier called MiMo-V2.6-Pro-UltraSpeed designed for latency-sensitive workloads.
Where It Ranks
According to VentureBeat’s reporting, MiMo-V2.6-Pro scores 46 on the Artificial Analysis composite Intelligence Index, placing it first among open-weight models and tied with the concurrently released Grok 4.7. That puts it ahead of Kimi K3, Qwen 3.8 Max, DeepSeek V4.1 Flash (39), and DeepSeek V4.1 Pro (36).
On the agentic software engineering benchmark DeepSWE v1.1, Pro scores 71.9% and Flash scores 67.9%. For context, GPT-5.6 Sol max scores 72.7% and Claude Fable 5.1 max scores 70.0%, placing Pro within 1 percentage point of the leading closed-source models on this task-completion measure.
| Model | Params (total / active) | Context | API price (input / output per 1M) | DeepSWE v1.1 |
|---|---|---|---|---|
| MiMo-V2.6-Pro | 1.02T / 42B | 1M tokens | $0.435 / $0.87 | 71.9% |
| MiMo-V2.6-Flash | 309B / 15B | 1M tokens | $0.14 / $0.28 | 67.9% |
| Grok 4.7 | 2.1T / N/A | 500K tokens | $2 / $6 | 71.0% |
| Claude Fable 5.1 (max) | closed | 200K tokens | approx $15 / $75 | 70.0% |
| GPT-5.6 Sol (max) | closed | 200K tokens | approx $15 / $60 | 72.7% |
Sources: Xiaomi official launch page, Artificial Analysis Intelligence Index, VentureBeat.
How Xiaomi Trained It
Xiaomi describes the V2.6 training approach as “You Only RL Once”: a single reinforcement learning run covering coding, general agent tasks, visual reasoning, and cybersecurity simultaneously rather than separate per-domain training passes. According to RuntimeWire and temperature2.com, the combined training cost was approximately $3.47 million for roughly 750,000 mixed RL trajectories, completed in fewer than six days.
Rather than relying on simple pass/fail scoring, the models learned by comparing multiple successful pathways to the same outcome, a technique Xiaomi says allows capabilities to reinforce and generalize across domains not explicitly represented in training.
Xiaomi has also open-sourced the training infrastructure:
- 7,000 RL task environments
- An end-to-end RL framework
- Composable mini-harnesses
- The full technical report
For enterprise teams building evaluation infrastructure, the environment library is relevant beyond the model weights themselves. It lets organizations assess whether the same RL approach transfers to their own task distributions, which is increasingly useful as enterprises define domain-specific reliability benchmarks for production agent systems.
Why It Matters for Enterprise AI
The Open-Weight Gap Has Narrowed Significantly
For most of 2026, the gap between closed-source frontier models and the best open-weight alternatives was wide enough that enterprise teams defaulting to GPT-5.6 Sol or Claude Fable 5.1 for complex agent workloads had an easy justification. MiMo-V2.6-Pro’s 71.9% on DeepSWE is close enough to the proprietary leaders to force a real evaluation.
For a team running thousands of agent calls per day on coding tasks, a model that performs within 1 percentage point of a proprietary frontier model at roughly 1.4% of the API cost changes the economics considerably. The Flash variant at $0.28 per million output tokens closes the performance gap further than the price difference would suggest, scoring 67.9% on DeepSWE while trailing Pro by just 4 points across several other agent benchmarks.
MIT License Changes the Compliance Calculation
Enterprise teams frequently cite licensing as the constraint that prevents adopting otherwise capable Chinese open-weight models. MIT license terms change that calculation. Unlike models with non-commercial or restricted licenses, MiMo-V2.6 permits commercial deployment, fine-tuning on proprietary data, and redistribution without revenue caps or research-only restrictions.
For regulated industries or organizations with data sovereignty requirements, the self-hosting option matters as much as the license terms themselves. A model that runs on your own infrastructure, with no external API call leaving your network, removes the compliance friction that stops many enterprise AI deployments early. Kimi K3 reaching Amazon Bedrock addressed managed cloud distribution, but on-premises self-hosting requires MIT-level licensing and available weights, which MiMo-V2.6 now provides at the top of the open-weight performance tier.
Flash for Volume, Pro for Complexity
Xiaomi’s two-tier positioning maps onto a real enterprise routing pattern. Most enterprise agent architectures do not select a single model for all tasks: simpler, high-frequency calls go to cost-efficient models, and complex, long-horizon tasks get the more capable option.
MiMo-V2.6-Flash fits the first category with unusual strength. At $0.28 per million output tokens and 67.9% on DeepSWE, it sits well above the commodity tier on performance while staying near the bottom of the pricing tier. According to Xiaomi’s own benchmarks, Flash stays within a few points of Pro across AutomationBench (52.3% vs 53.1%), Terminal Bench 2.1 (87.6% vs 89.9%), and MiMo Visual Coding (71.5% vs 72.3%).
For workloads where an enterprise might launch thousands or millions of agent runs daily, a $0.87 vs $0.28 output cost difference compounds quickly. At 10 million output tokens per day, the difference between Pro and Flash is roughly $5,900 per day, or over $2 million per year.
What It Does Not Change
MiMo-V2.6-Pro does not beat proprietary frontier models on every measure. Gaps remain in terminal-intensive operations and certain cybersecurity benchmarks, and the performance figures in this article are Xiaomi’s own reported numbers, not independently reproduced results.
The Artificial Analysis composite score of 46, while topping the open-weight list, sits below the scores for Claude Fable 5.1 and GPT-6 Astra on that same ranking. “First among open-weight models” is not “best model available overall,” and any enterprise evaluation should include live testing on domain-specific workloads before treating leaderboard position as a deployment decision.
What Comes Next
The open-source infrastructure release is arguably as significant as the model weights. Xiaomi publishing 7,000 RL environments and the full RL framework means external teams can evaluate, extend, and build on the training approach. If the “You Only RL Once” cross-domain mixing proves transferable, future open-weight models built on this infrastructure could close the remaining gap to proprietary frontier models further.
For enterprise AI teams currently running cost benchmarks or planning agent infrastructure, MiMo-V2.6 belongs in the evaluation set. For teams already comfortable with open-weight deployment, the Flash variant in particular warrants serious consideration as a high-volume agent backbone.
Enera helps enterprise teams build agentic AI systems with the right model selection, routing, and governance layer for their specific workloads. If you want to evaluate how open-weight options like MiMo-V2.6 fit your agent stack, book a call with the team.