--- title: "Xiaomi ships MiMo-V2.6 as a native omnimodal model and open-sources it" date: 2026-10-25 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-xiaomi-mimo-v26 language: en --- # Xiaomi ships MiMo-V2.6 as a native omnimodal model and open-sources it > Xiaomi's MiMo-V2.6 series arrives as natively omnimodal open-weight models, with a phone-maker pitching 'built in public' as its answer to closed frontier labs. A person points a phone camera at a broken dishwasher, speaks a sentence in Mandarin, and expects the assistant to see the part, hear the complaint, and write a working repair plan. That is the bar a modern assistant has to clear — and Xiaomi, a company best known for hardware, is now building the model to clear it in the open. ## Key takeaways - Xiaomi released and open-sourced the **MiMo-V2.6** series: a flagship **Pro**, a cost-balanced **Flash**, and a **Pro-UltraSpeed** variant promising up to 20x faster output at the same quality. - The models are described as **natively omnimodal**, handling text, image, video, and audio rather than bolting modalities together after the fact. - Xiaomi frames the work as the "RSI path" — scaling reinforcement-learning compute on verifiable, complex tasks so the model expands its frontier through exploration and feedback. - On company-reported coding and agent benchmarks, MiMo-V2.6-Pro lands just behind the strongest closed models and ahead of several open rivals. - The move continues Xiaomi's open-weight streak: the earlier V2.5 line shipped under an MIT license with a 1-million-token context and a 100-trillion-token community incentive plan. ## A hardware company's model strategy Xiaomi is not a lab that happens to make phones; it is a device maker that needs models to live on its phones, earbuds, cars, and glasses. That constraint shapes the product. An omnimodal model that can take a photo, a voice note, and a typed instruction in one turn is directly useful for a company whose surface area is physical objects, not web forms. The "built in public" language matters too. Xiaomi has been releasing weights, benchmarks, and adaptation code rather than hoarding a closed API, betting that a developer ecosystem beats a gated moat for a company whose strength is manufacturing scale. ## What "natively omnimodal" buys you Most multimodal systems train a text model and then glue on an image encoder or a speech module. Xiaomi says MiMo-V2.6 is omnimodal from the ground up, meaning the same backbone is expected to reason across text, image, video, and audio without a translation layer between them. The practical promise is fewer dropped contexts when a task mixes a screenshot, a spoken question, and a written follow-up — exactly the kind of mess real users generate. The pitch is less "our model is smarter" and more "our model fits the way people actually interact with devices." ## The numbers Xiaomi is showing The company's own benchmark tables put MiMo-V2.6-Pro close to the front of the open field. On DeepSWE v1.1, a software-engineering benchmark, the Pro scores 71.9 and Flash 67.9, trailing only the very top closed models. On GDPVal 2.1 (Artificial Analysis), Pro reaches 1673 against a leading closed model's 1735. On Toolathlon-verified, an agent-tool-use test, Pro scores 76.9. Other reported results include Automation Bench (Pro 53.1), ALE (Pro 31.6), and Terminal Bench (Pro 34.9). Read these as Xiaomi's claims, not independent verdicts. They are useful as a relative signal: the gap between a serious open model and the closed frontier has narrowed to single digits on several agent and coding tasks. ## Continuity from V2.5 The V2.6 release sits on top of a clear open-weight lineage. The MiMo-V2.5 series, open-sourced in spring 2026 under an MIT license, supported a 1-million-token context and ranked first among open-source models on GDPVal-AA and ClawEval at the time. Xiaomi paired it with the "Orbit" program: a 100-trillion-token community incentive, where the top approved tier was worth about 1.6 billion credits, or roughly 659 yuan (≈ US$93 / HK$720). That combination — permissive license, long context, and a token giveaway — is a recognizable playbook for seeding adoption. Developers are more likely to build on a model they can commercialize for free and test without a bill. ## Why a maker is a credible model builder now Five years ago, a phone brand shipping a frontier-class model would have sounded like marketing. Today the logic is concrete: device companies sit on enormous, modality-rich interaction data, they control the silicon path (Xiaomi's own chips and tight ties to suppliers), and they have a distribution surface no pure-play lab can match. When the model and the gadget are the same company's, latency, privacy, and offline use stop being roadblocks and start being features. ## Honest limitations - All benchmark figures here originate from Xiaomi's own release materials. We did not run MiMo-V2.6 independently, and the scores should be read as the company's positioning, not a neutral ranking. - "Omnimodal" claims are easier to demonstrate in a demo than to hold up across noisy, real-world inputs; robustness on mixed-modality tasks under load is unverified by third parties. - Open weights under MIT do not guarantee frictionless deployment on every device class; on-device performance depends heavily on quantization and hardware support. - The 100-trillion-token Orbit plan was a time-limited campaign tied to the V2.5 launch; treat it as historical context, not a standing offer. - Cross-lingual and cultural performance beyond Chinese and English is not thoroughly documented in the materials we reviewed. ## What readers can do now - **Builders:** grab the MIT-licensed weights and prototype a mixed-modality feature (photo + voice + text) where a single-turn assistant genuinely helps. - **Businesses:** benchmark MiMo-V2.6 against your current closed model on one real agent task before assuming the premium API is worth the bill. - **Observers:** watch whether device-makers, not labs, become the default home for omnimodal assistants — distribution beats research prestige when the model lives in your pocket. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-xiaomi-mimo-v26 Free to quote with attribution and a link to the original.