Xiaomi MiMo-V2.6 has launched as two native multimodal models—Pro and Flash—with open weights, a technical report and reinforcement-learning resources. Xiaomi disclosed the release on September 22 after six days of publicly tracked training and kept API pricing aligned with its earlier V2.5 series.
Xiaomi MiMo-V2.6 makes the training stack part of the launch
Xiaomi says the two models completed 30 reinforcement-learning steps each across about 750,000 trajectories. It reports training costs of roughly $2.62 million for Pro and $850,000 for Flash, plus gains on the DeepSWE v1.1 software-engineering evaluation. These figures are from the company’s own disclosure; VentureBeat and ITHome independently reported the release and its open-weight positioning.
Xiaomi MiMo-V2.6 is notable not only for two new models but for releasing the surrounding reinforcement-learning environments and code, giving researchers more material to inspect than a closed API benchmark alone. The open package includes more than 7,000 task environments spanning software engineering, vulnerability reproduction, knowledge work and web design.
| Release | Role | Published training cost |
|---|---|---|
| MiMo-V2.6 Pro | Higher-capability model | About $2.62 million |
| MiMo-V2.6 Flash | Lighter model | About $850,000 |
| Distill-Qwen-9B | Research starting point | Not separately stated |
What developers can actually use
The company says both models support computer use, office tools, information retrieval and code-driven content creation. It also launched a desktop client and API access, while the open collection includes a distilled 9-billion-parameter starting model and a composable harness. The UltraSpeed Pro option is advertised at up to 20 times the regular inference speed, but teams should validate latency and output quality on their own hardware and region.
For India’s AI engineering ecosystem, the useful opening is reproducibility: smaller labs can inspect the task environments and adapt the training recipe instead of relying only on an endpoint. The limitation is equally clear—open weights do not automatically reveal all training data or make vendor benchmarks neutral. Readers can compare the release with the NASA–IBM open foundation model and Lapaas Voice’s look at measuring AI-agent costs.
A useful evaluation should run Pro and Flash on the same multilingual coding, document and computer-use tasks. Teams should track accepted completions, inference latency, memory demand and human correction time rather than adopting the published index as a purchasing shortcut. Researchers reproducing the reinforcement-learning results should pin code versions and document any unavailable data. That makes the open artifacts a starting point for verification instead of a substitute for it.
The two-model design can support routing rather than a winner-takes-all choice. A team might send routine extraction and drafting to Flash, then reserve Pro for tasks that fail an initial check or require deeper reasoning. That architecture only saves money when routing quality is measured and retry loops remain bounded. Published API rates and benchmark scores should therefore feed a routing experiment, not determine the production policy by themselves.
Frequently asked questions
What is Xiaomi MiMo-V2.6?
It is a two-model multimodal AI release comprising Pro and Flash, with API access and openly released weights and research resources.
Is Xiaomi MiMo-V2.6 open source?
Xiaomi has released weights, a technical report, task environments and reinforcement-learning code. Users still need to review each repository’s licence and deployment terms.
Why are there Pro and Flash versions?
The split gives developers a capability-oriented option and a lighter option, allowing cost and latency trade-offs to be tested against the same family.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
