# MiMo-V2.6 Is MIT-Licensed. I Would Test the API First. **Summary:** Xiaomi released MIT-licensed MiMo-V2.6 weights, but the official Pro serving recipe assumes serious infrastructure. I would test the low-cost API before pricing a self-hosted cluster. - Canonical: https://markhuang.ai/news/mimo-v2-6-open-weights-api-first - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-09-21 - Section: News - Tags: Xiaomi MiMo, Open Weights, AI Infrastructure, Model Deployment, AI Agents - Source: [Xiaomi MiMo](https://mimo.xiaomi.com/mimo-v2-6) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![An open, liquid-cooled AI compute system fills a data center while a single cable connects it to a small access terminal](https://cdn.markhuang.ai/news/mimo-v2-6-open-weights-api-first/hero.webp) *MiMo-V2.6 is open enough to inspect and operate. The machine behind that freedom is still large.* [Xiaomi's MiMo-V2.6 release page](https://mimo.xiaomi.com/mimo-v2-6) turns last week's public training run into two downloadable, MIT-licensed models. MiMo-V2.6-Pro has 1.02 trillion total parameters with 42 billion active, while Flash has 309 billion total and 15 billion active. Both accept text, images, video, and audio, and Xiaomi lists a 1 million-token context window. This fills in much of what I wanted to see. I was skeptical when Xiaomi first showed the training bill without the finished recipe. Now I can inspect the weights, model cards, architecture, deployment commands, technical report, and hosted endpoints. I would still start with the API. The open release is real, but operating it is a separate decision. ## Xiaomi closed the artifact gap On September 16, I wrote that [the live dashboard made MiMo-V2.6's bill clearer than its recipe](https://markhuang.ai/news/mimo-2-6-live-training-bill-not-recipe). At that point, I could see costs, restarts, and checkpoint movement, but not a model card, license, architecture, or downloadable weights. The release resolves most of that complaint. The [Pro model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) and [Flash model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) both carry the MIT license. They spell out each mixture-of-experts architecture, activated parameter count, multimodal encoders, context length, evaluation table, and serving examples. I can work with that. A promise to open the model later would not be enough. Xiaomi also says it is publishing the technical report, training environments, and reinforcement-learning code. I have not reproduced the training run, and publishing components does not prove that an outside team can regenerate the checkpoints. Still, the release gives researchers and operators concrete material to inspect. I no longer think "where are the artifacts?" is the right objection. ## The operating bill survives the license change The Pro deployment example is the part I would read before downloading anything. Xiaomi's official SGLang command uses tensor parallelism of 16, data parallelism of 2, expert parallelism of 16, and two nodes. Its vLLM example uses tensor parallelism of 8. Those are example configurations, not stated minimum hardware requirements, but neither looks like a casual workstation project. Flash lowers the scale from 1.02 trillion total parameters to 309 billion. That is a serious reduction, yet Flash remains a large model. The weights matter for teams that need control over deployment, modification, or where data is processed. A developer who mainly wants to learn whether MiMo is good at a real workload would turn the first test into an infrastructure exercise by self-hosting it. > **Info:** > > I would test Flash and Pro through the hosted API, keep the same prompts and acceptance checks, then price self-hosting only if control, sustained utilization, or data policy makes it necessary. The hosted prices make that order hard to ignore. Xiaomi lists MiMo-V2.6-Pro at $0.435 per million uncached input tokens and $0.87 per million output tokens. Flash costs $0.14 and $0.28 respectively. Cache-hit input is cheaper still. Those rates do not include my review time or the cost of failed agent runs, but they make a pilot cheap enough that I do not need a hardware thesis before I have a quality result. ## The benchmark number changes with the evaluation Xiaomi says each model completed 30 reinforcement-learning steps over roughly 750,000 trajectories. It reports that the Flash run cost about $850,000 and the Pro run about $2.62 million. On the live-run DeepSWE v1.1 evaluation, Xiaomi says Flash rose from 48.8 to 65.68 and Pro from 58.4 to 72.57. The released model cards list 67.9 for Flash and 71.9 for Pro on DeepSWE v1.1. That does not make either table false. It does mean I should not collapse the checkpoint dashboard and the release evaluation into one portable score. The model card also shows substantial variation by task: Pro scores 89.9 on Terminal Bench 2.1 and 34.9 on Terminal Bench 4.0 under Xiaomi's reported setups. This is where the launch claim that Pro is the strongest open-source model loses me. Xiaomi bases it on a 46.32 Artificial Analysis Intelligence Index score. A composite index can place the model in the field. It cannot tell me whether MiMo will finish a migration cleanly, call my tools correctly, or leave a reviewer with a costly repair job. Public reaction is still too early to call a verdict. A [Hacker News exchange about the live run](https://news.ycombinator.com/item?id=49732850) had to clarify that the six-day event was reinforcement-learning post-training, not training the whole model from scratch. A new [LocalLLaMA thread on the Pro weights](https://www.reddit.com/r/LocalLLaMA/comments/1wmntwv/xiaomimimomimov26prorl_hugging_face/) was mostly initial interest. Independent workload reports will matter more than that first burst of attention. ## I would buy evidence before hardware My first test would use the cheaper Flash endpoint on a fixed set of repository tasks with known acceptance criteria. I would record completed tasks, repair turns, latency, token cost, and review time. Then I would run the misses through Pro. That comparison answers two immediate questions: whether the larger model changes the outcome, and whether it saves enough human time to justify its higher price. Only after that would I evaluate local deployment. A company with steady demand and suitable infrastructure may reach a different cost equation, especially if its inputs need to stay inside its own boundary. The MIT license and downloadable weights make that choice possible. The cluster still has to be paid for, operated, and kept busy. MiMo-V2.6 is a better release than the live dashboard alone led me to expect. Xiaomi followed a public, messy post-training run with actual artifacts. That earns credit. For most teams, the first move should still be a small API experiment that shows whether the model deserves a machine of its own.