On September 21, 2026, Xiaomi, the company most people know for phones, put two large AI models on the internet for anyone to download. MiMo-V2.6-Pro and MiMo-V2.6-Flash are released under the MIT licence, which is about as permissive as software licences get. You can run them on your own machines, change them, build a product on them and sell it, without asking Xiaomi or paying it a cent. That is what "open weights" means, and it is the reason this release matters more than one more model on a leaderboard. The open-source AI tier is now crowded with serious Chinese labs, the closed frontier is still American, and the gap between them has become the most useful number in the industry to understand.
What Xiaomi Released, and Why a Phone Maker Bothers
The facts come from Xiaomi's own release, summarised by SiliconANGLE on September 22, 2026. MiMo-V2.6-Pro has about 1.02 trillion parameters in total but only about 42 billion "active" on any given token, a mixture-of-experts design where the model routes each piece of text to a small slice of itself. Flash is the same idea smaller: about 309 billion total, about 15 billion active. Both take text, images, video and audio natively, and both handle a context of about one million tokens, roughly a few thousand pages in one prompt.
Why would a hardware company give this away? Because for Xiaomi the model is not the product. Phones, cars, appliances and the services on them are the product, and an assistant that runs across all of them is the entry point. A capable model of its own removes a supplier from the bill of materials, lets the company shape the assistant to its devices, and, if developers adopt the weights, makes Xiaomi's stack a place other people build. Xiaomi also published the training report, more than 7,000 agent-training environments and a distilled 9-billion-parameter model for research, which is the kind of thing you release when you want a community, not just a download count.
It is fair to say that plainly, and also fair to keep the scepticism: this is a strategic move by a company with a phone business to protect, not philanthropy. Both things are true, and the second does not make the weights any less useful.
The Open-Source AI Field in September 2026
Six names come up in almost every serious comparison, and they are not interchangeable. Using Artificial Analysis's Intelligence Index v4.3 as one common yardstick, with the caveat that every leaderboard measures something slightly different:
Zhipu's GLM-5.3 and Moonshot's Kimi K3 lead the confirmed open tier at 44, GLM as the coding-focused workhorse, Kimi with native multimodality and a million-token context. Alibaba's Qwen3.8 is the broadest family: a 2.4-trillion-parameter sparse flagship scoring 40, plus a 27-billion multimodal checkpoint under Apache 2.0 that runs on far cheaper hardware. DeepSeek V4 Pro scores only 36 on the composite. By the reported numbers it is the strongest open model at real software engineering. Meta's Llama 4 is still widely deployed, though Meta has since moved its newest flagship, Muse Spark, closed. And now MiMo-V2.6, with a vendor-reported 46.32 that, if the leaderboard confirms it, would put a phone maker at the top of open weights.
Treat every one of those numbers as a snapshot. The index changed its own composition between v4.2 and v4.3, several labs ship a new checkpoint every few weeks, and a two-point gap is inside the noise of a new release. What the snapshot does show reliably is the shape of the field: the open tier is deep, it moves fast, and the companies driving it are mostly Chinese.
Open vs Closed: The Gap Is Real, and It Is Shrinking Unevenly
On the composite index the closed frontier still wins clearly. Claude Fable 5.1 and GPT-6 Astra sit at 53; the best confirmed open models at 44. Nine points on this scale is not cosmetic. It shows up as fewer mistakes on long, messy, multi-step work, the kind of task where a model has to hold a plan for an hour and not drift.
Look at individual tasks and the picture splits. On knowledge, maths and graduate-level science questions, open models now match or beat closed ones, as several 2026 comparisons note. On production coding, DeepSeek V4 Pro's reported 80.6% on SWE-bench Verified is within about a point and a half of Claude Opus. Where closed models keep a clear lead is human preference in open-ended chat and the hardest agentic workflows.
A homely way to hold this: think of a good regional airline and the flag carrier. On most routes you will not notice the difference, and the regional one is a fraction of the fare. On the longest, most complicated itinerary, with three connections and a storm, the flag carrier still gets you there more often. The gap has narrowed every year. It has not closed, and on the hardest trips it still matters.
Then there is the advantage no benchmark measures. An open-weight model can be run inside your own network, next to your own data, on hardware you control. You can read what it is, fine-tune it on your documents, and keep it running if the vendor changes its prices or its terms. For a hospital, a law firm, a factory or a government department, that is often the deciding factor, and it is why the score comparison is the beginning of the decision rather than the end.
Why Open Weights Matter to People Who Will Never Train a Model
The obvious beneficiary is the individual developer. A solo builder can download a 27-billion-parameter Qwen or a distilled MiMo, run it on a single rented GPU or a well-equipped laptop, and ship a product with no per-token bill and no API key to leak. The cost of trying an idea drops from a monthly invoice to an afternoon.
The less obvious beneficiary is the small business. When we build AI agents for clients, the question "can this run inside our own systems?" comes up in most engagements, usually from the owner rather than the IT person. A logistics firm does not want its order data leaving the building; a clinic cannot send patient notes to an external API at all. Open weights turn that from a blocker into a configuration choice, and the MIT licence on MiMo and DeepSeek means the legal review takes a morning, not a quarter.
The third beneficiary is the university lab or the student. Xiaomi's release includes the training infrastructure and those 7,000-plus reinforcement-learning environments. A research group that could never afford to train a frontier model from scratch can now study how one was trained, reproduce parts of it, and publish. That is how a field gets deeper rather than just bigger.
Underneath all three is the same structural point: open weights break single-vendor lock-in. The 2023 pattern, where every AI feature in the world was a thin wrapper around one company's API, has a second option now. That alone changes negotiating power for every buyer, whether they ever self-host or not.
Why the Closed Frontier Is American and the Open Frontier Is Chinese
This is the part of the story that gets flattened into slogans, so it deserves mechanisms instead. Five forces push in the same direction, and each is a trend with exceptions, not a law.
Compute and export controls. Access to top-tier training chips in China has been constrained and, worse for planning, unpredictable. The Congressional Research Service documents the framework; in early 2026 CNBC reported that even approved H20 and H200 sales had produced no shipments. If you cannot count on the biggest chips, you optimise for efficiency: sparse mixture-of-experts designs like MiMo's 1.02 trillion total with 42 billion active, and DeepSeek's famously frugal training. And if you cannot serve the whole world's inference from your own data centres, publishing the weights lets everyone else's hardware do it for you.
Capital structure and business model. OpenAI and Anthropic sell the model itself, through subscriptions and APIs, so the weights are the crown jewels. Alibaba, Xiaomi, Moonshot and Zhipu sit inside or beside cloud, phone and e-commerce businesses that make money on the ecosystem around a model. For them, open weights are a customer-acquisition cost that pays back in cloud consumption, device sales and developer loyalty. Same technology, opposite incentive.
Talent and organisation. The American frontier labs are research organisations with the culture, and the funding, to run multi-year closed programmes. Many of the Chinese labs grew out of engineering-heavy internet companies with a strong open-source tradition and a habit of shipping in public, fast, and iterating on feedback. Releasing weights every few weeks is a delivery style, not just a licensing decision.
Policy and industrial direction. In the US the governing narrative around frontier models is safety, controlled access and competitive advantage, which favours closed release. In China the priorities include domestic substitution and a fierce competition among cloud providers to become the platform developers standardise on, which favours giving the base layer away.
Market strategy. When the incumbents' APIs already dominate, the fastest way to win developer mindshare is to be the thing developers can run themselves. Open weights set standards, fill tutorials and forum answers, and make a lab's architecture the default in a thousand small projects. Monetisation happens above that layer: enterprise versions, cloud instances, devices.
Now the exceptions, because they are real. The US ships strong open weights: Google's Gemma 4 under Apache 2.0, OpenAI's gpt-oss line, NVIDIA's Nemotron models. China has closed products, and Meta, the company that made "open" a strategy with Llama, has moved its newest flagship closed. Read the pattern as a tendency produced by incentives, and expect it to bend when the incentives do.
What to Actually Do With This
If you are an individual developer: pick two open models of different sizes, say a 27-billion Qwen or Gemma 4 for your laptop and MiMo-V2.6-Flash or DeepSeek V4 on a rented GPU, and run your own ten-task test on the work you actually do. Ignore the composite score. A model that is two points worse on the index but nails your document format is the better model for you.
If you run a company: self-host an open model when the data must stay inside, when volume is high enough that per-token pricing hurts, or when you need to fine-tune on your own material. Keep buying the closed API when the task is at the frontier of difficulty, when you have no one to run infrastructure, or when you are still finding out what the workload is. Most of the businesses we work with end up with both, and that is the right answer, not a failure to decide.
Whichever way you go, three checks before anything goes live. Read the licence, not the headline: MIT and Apache 2.0 are simple; Llama's terms and some Qwen and Kimi conditions have clauses for large services and, in Llama's case, EU users. Decide where the data goes and who can see the logs. And write down the real cost of self-hosting, including the person who will keep it running at 2 a.m., before comparing it to an API bill. We covered the human side of that equation in what "AI does 26% of the work" really measures, and we run AI training programmes for teams making exactly these calls.
The most useful thing about MiMo-V2.6 is not that Xiaomi has joined a race. It is that the race now has enough serious entrants that no one has to pick a side. That is new. Choose the tool that fits the job, the data and the budget, and change it when a better one arrives. That is a healthier position than any of us were in three years ago. If you want help deciding for your own business, talk to us.
Further reading, by search term: Artificial Analysis Intelligence Index v4.3; MiMo-V2.6 technical report; Qwen3.8 licence; DeepSeek V4 SWE-bench Verified; US semiconductor export controls 2026.
