The best model anyone can download changed hands
Edition of 2026-09-22
# AI For All ### Tuesday, September 22, 2026
As of today, a lawyer, an accountant or a small advisory firm can download a model that reads text, images and on-screen operations and run it on hardware it controls, because the strongest openly downloadable model in the world was published with its weights attached.
The best model anyone can download is no longer American or European
On Tuesday, Xiaomi released and open-sourced a family of models called MiMo-V2.6. Two of them are large and fully multimodal, named Pro and Flash, and a third, smaller one is aimed at researchers. The independent Artificial Analysis Intelligence Index scores the Pro model at *46 points*, which places it above Kimi K3 and Qwen3.8 Max, and makes it the strongest model that anyone can simply download and run privately.
What makes it a story rather than a score is what came with the weights. Xiaomi trained these models in public for less than *6 days, published the compute bill for the larger run at 2.62 million US dollars*, and released the technical report, the training environment and the reinforcement learning code alongside the model itself. The price of using it through Xiaomi's API is unchanged from the previous generation.
Two reserves, and the first one matters if you intend to build on it. The ranking is a third-party measurement and that measurement exists, Artificial Analysis published its own entry for the model on September 21, but the sentence calling it the most powerful open model available is Xiaomi's own. The second: nobody here has checked the licence attached to the weights. Read it before you put it inside a business.
Read the release at Xiaomi MiMo.
A price per unit is not a price per job
SpaceXAI published Grok 4.7 on Monday, and its announcement leads with cost. The rate is *$2/$6* per million units billed, identical to the model it replaces, and far below the two flagships the announcement compares itself to. Then an independent lab measured what the model consumes to finish the same work, and the picture changes.
The evaluation found that the gains come with much heavier usage. Grok 4.7 spends *125%* more billed output than the model it replaces to complete the same tasks, and more still against one rival's flagship. That is the number to take away. A price per unit is not a price per job, and these two have just come apart. If you budget by usage rather than by subscription, your cost per task moved in the opposite direction to the headline price.
The same evaluation is not unkind. It puts the new model into the top four labs on its index, and fourth among coding agents, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5. It also records two tasks where the model got worse. A strong release with a real bill attached.
Read the independent evaluation at Artificial Analysis.
The laptop that hides a subscription inside the price
Google opened pre-orders on Monday for Googlebook, a laptop line it first showed in May, and the interesting part is not the hardware. Every machine in the range includes *12 months* of Google AI Pro, the paid tier of its assistant, with five terabytes of storage. The assistant is built into the desktop rather than bolted on: it reads whatever is on screen when asked, turns speech into formatted notes, and generates small tools from a plain description.
This is a serious supply chain, not a single-vendor experiment. Five established manufacturers are building the machines, Acer, ASUS, Dell, HP and Lenovo, on Intel and Qualcomm processors. Google also promises updates for up to *10 years*, which is far longer than the three to five years most companies keep a laptop, and that changes the arithmetic on the subscription bundled inside the price.
The reserves are the ones a buyer should hold on to. You cannot use it today, you can only reserve it today, and it reaches shelves on October 4 in the United States and October 5 in six other countries. The assistant features are described as what the laptop is designed for, not as something anyone outside Google has measured yet.
Read the announcement at Google.
Weak signal
Nine mathematicians announced a new body on Monday, the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton. The members include Edward Witten, Timothy Gowers and Martin Hairer. Their first job is advising OpenAI on how to release a large number of mathematical results that the company says its own model produced.
The group says it is independent of any AI company, that members are not paid for the work, and that it will publish its recommendations. It came together after OpenAI approached some of the members about forming an external advisory board, and the members agreed with the company to make the body independent instead. Nothing here changes what a reader can do today. It changes who gets asked before a claim in mathematics is published, and that is worth watching.
The next thing to watch is the bill. If the best model anyone can download now matches the paid flagships, the price of intelligence stops being the story, and the cost of running it becomes the story.