The open-model tier stopped trailing today, and it did so from an unexpected direction. Xiaomi, of all companies, dropped MiMo-V2.6 under an MIT licence and claims the highest score of any open model, arriving on X overnight before most outlets had noticed. Alongside it, xAI shipped Grok 4.7 as a same-price, same-speed upgrade tuned for the hours-long coding sessions that separate a useful agent from a toy, and Perplexity folded two video generators into Computer so a campaign clip now lands in the same thread as the copy. Qwen's weekend image model finally got the serving support that makes it runnable, and OpenAI broadened its free Academy into role-based tracks. Nobody at the closed frontier shipped a new model, and the day's verdict is that the interesting competition has moved to price, licence and how few tools you need open at once.
MiMo-V2.6 Pro and Flash arrive as MIT-licensed open weights, with Pro claiming the top open spot
This is the most consequential open weights drop of the day and it landed on X overnight before most outlets had written it up. Pro is a sparse mixture of experts model with 1.02 trillion total parameters, 42 billion active per token and a one million token context window, and it scores 46 on the Artificial Analysis Intelligence Index, which Xiaomi says is the highest of any open model. Flash is the lighter sibling at 309 billion total and 15 billion activated per token, for people who want something they can realistically serve.
The licence matters as much as the benchmark: MIT means you can self host, fine tune and ship commercially without asking anyone, and Xiaomi has published the technical report, the reinforcement learning environments and the training code alongside the weights. It is the sharp counterpoint to the research-only drift in other Chinese open weights, and if you have been holding off on moving workloads off a closed API because the open options were a tier behind, this closes most of that gap. It is already live on OpenRouter, in MiMo Code and through Xiaomi's own API platform.
Grok 4.7 lands as a same-price, same-speed upgrade aimed at long coding sessions
If you already pay for Grok through the API or use it inside Cursor, this is a free capability bump rather than a new line item, because SpaceXAI is serving 4.7 at the same price and speed as 4.6, at two dollars per million input tokens and six dollars per million output under 200k context. The interesting part is what it was tuned for: a longer reinforcement learning run weighted towards problems that take hours rather than minutes, which is the difference between a model that writes a function and one that can hold a refactor together. It posts 46.3 per cent on CursorBench 4.0, 71.0 per cent on DeepSWE and 38.0 per cent on Terminal-Bench, putting it at the price and performance frontier rather than the raw capability frontier.
Safeguards were rebuilt too, with 62.4 per cent on LatchBio's biosafety benchmark and only 3.3 per cent of risky dual use prompts allowed through on HackerBench v0.3, while still permitting legitimate security work. Access is live now through the Grok API, Grok Build, Cursor, third party coding harnesses and the usual model routers.
Perplexity Computer can now generate video in the same thread as the copy
Perplexity has wired MiniMax H3 and ByteDance Seedance 2.5 into Computer, so a request for a campaign clip, product demo or social asset now returns finished video sitting next to the written copy and the stills in one conversation. The workflow change is the point rather than the models themselves, since both are already available elsewhere: the value is not having to export a brief into a separate video tool and then reconcile the output with everything else.
For anyone running small-scale marketing production, that removes a genuine chunk of the busywork between idea and asset. It is live now for Pro subscribers at twenty dollars a month and Max at two hundred, with Max getting the higher rate limits. Coming days after the effort slider, it is worth testing before you renew a standalone video subscription, because for a lot of light users this will simply absorb that spend.
Qwen-Image-2.1 gets day-zero serving support, which is what makes it usable
The model weights themselves went out on 20 September, just outside this window, but the announcements that landed inside it are the ones that determine whether anyone can run it. Qwen confirmed SGLang-Diffusion now serves Qwen-Image-2.1 for text to image, multi-image editing and transparent RGBA output, hitting 1024 by 1024 generation in 18.7 seconds and image editing in 21.7 seconds on a single RTX 4090 with CPU offload. A browser-based demo also went up on Hugging Face Spaces, so you can try the model without touching a GPU.
That combination matters because Qwen-Image-2.1 is a 7 billion parameter model designed to run on consumer hardware, and a release of that shape is worth very little until the serving stack catches up. One caveat before you build on it: the weights carry a research licence that bars commercial use without a separate grant, so check that before it goes anywhere near client work.
OpenAI Academy adds role-based learning paths for developers, leaders, educators and students
OpenAI has broadened its free training programme from a single general course into separate tracks aimed at different jobs, joining the existing Apply AI at Work pathway. The framing is deliberately practical: learners work on real tasks, give instructions, add context and review results, and there is an assessment at the end that earns a course badge. Whether a badge from a model provider carries weight with an employer is an open question, but the courses cost nothing and the agent-focused material is more current than most paid alternatives.
If you are trying to get a team past the copy-and-paste stage of using AI, this is a reasonable shared starting point rather than writing your own internal training. No product or model changes ride along with it.
Industry themes
Two of the three biggest items today are open weights releases, and the Xiaomi one surfaced on the company's own X account hours before mainstream coverage caught up. That is the pattern worth noting: the open model tier is no longer trailing a generation behind, and the announcements now arrive through company channels first, so anyone relying on press coverage alone is reading yesterday's news.
The second theme is that price competition has replaced capability claims as the headline. Grok 4.7 leads with the same price as its predecessor rather than a benchmark record, and Qwen-Image-2.1 is explicitly sized for a consumer GPU, which tells you where the pressure is. When a new model's launch line is about cost and hardware rather than a leaderboard, the market has matured.
Third, features are quietly being absorbed into bundles rather than sold separately, as Perplexity folding two video models into Computer shows. The practical consequence for anyone running a stack of subscriptions is the same each week now: audit the standalone tools before they renew, because the assistant you already pay for keeps swallowing them.