A model small enough to keep in the office
Edition of 2026-08-15
As of today, a small firm that cannot send client files to an outside service can run a model that reads its documents and its images, on one machine in its own office.
The model you can actually run
Qwen has published the weights of a compact model from its newest generation, under a permissive licence. Anyone can download it and run it privately. At 27B parameters it is a fraction of the trillion-scale open models Chinese labs have shipped all summer — the ones nobody outside a data centre could load. This one fits on hardware a business already owns, and the community had quantised builds up within a day.
It reads pictures and documents, not just typing: diagrams, scanned paperwork, hour-long video. It holds up to 1,000,000 tokens' worth of text in mind at once at the top setting, roughly a shelf of contracts. The hosted version through Qwen's own cloud is still described as coming soon, which makes the download the product today.
The comparison table on the page is Qwen's own, run on tests Qwen chose, and no independent measurer had evaluated this model when we looked. So take the size and the licence as fact, and the scores as marketing until somebody neutral checks them. Qwen model card
Claude's text now carries an invisible signature
Anthropic says future Claude models will watermark the text they write, everywhere, not just in Europe. The mechanism is subtle: when the model chooses between two words that would both be fine, a secret key settles the choice instead of chance. Nothing is added to the text, no hidden characters, no extra cost, and a reader sees no difference at all. Someone holding the key can then estimate whether Claude was involved.
The reason is regulatory. Anthropic signed the European transparency code alongside around 190 signatories, a figure the European Commission confirms on its own page. Anyone selling AI-written work into the EU market should assume this becomes the norm rather than one company's choice.
Then the limits, which Anthropic states itself and which matter more than the announcement. Short passages cannot be checked. Neither can code, or a passage of hard facts, where the right word is forced. If Claude only proofread your writing there is too little of Claude in it to detect, and a full rewrite removes the mark. The detection tool does not exist yet — an API is promised, undated. So the marking is live and the checking is not, which is a strange interval to leave open. Anthropic
A new flagship arrived inside a coding subscription
Z.ai released its new flagship model yesterday to everyone on its coding plan. The programming interface is not open — the documentation says it is coming soon — so the subscription is the only door today. Anyone already paying for it woke up to a better model at the same price.
What the company claims about the build is the interesting part. It did not train a new model. It took the same base as the previous version and taught it differently afterwards, and says that accounts for every gain. Nathan Lambert, who covers open models independently, puts it at 750B parameters, about a third the size of the largest Chinese open model this summer, and says the weights land publicly in two weeks.
Every performance number on the vendor's page is the vendor's own, measured on the vendor's own bench, so none of them appear here. Diary the fortnight instead: when the weights land, anyone can measure it for themselves. Z.ai documentation · Interconnects
What people applaud is not what people run
Hugging Face published its half-yearly read of the open model world, and one finding should change how you read AI news. Take the twenty-five most downloaded models this year and the twenty-five most liked. Exactly one appears on both lists. Not a single model released this year makes the download list.
The shape underneath is extreme. There are now 2.96 million public model repositories on the site, and 1.5% of them account for 99.2% of all downloads. One small, elderly text model was pulled 1.55 billion times in seven months while collecting a few thousand likes; this summer's celebrated frontier model was pulled about sixty times per like.
A like records a reaction. A download records a dependency — something wired into software that runs on a schedule, for years. This is the hub reporting on its own traffic, and automated jobs inflate download counts, which the report concedes. It is still the better guide to what a supplier decision should rest on. Hugging Face
Signal faible
Google released HEIR, an open-source compiler that converts an already-trained model so it can work on encrypted data it never unlocks, with demonstrations including credit card fraud detection and network intrusion monitoring, built with named hardware partners. Today it is a tool for cryptography engineers and Google says the speed benefits are still to come. The day a bank ships a service on it — your data processed without ever being visible to the processor — that becomes a story. The post is at blog.google, under security.
Watch the fortnight: when Z.ai's weights land, the claims stop being the company's to make.