Qwen3.8-27B
Your expense report sits behind a screen full of buttons. Qwen3.8-27B, from Alibaba's Qwen team, is built to look at a screen like that and click the right thing, the way you would. The whole model is a fifty-five gigabyte download, and anyone can take it.
Roughly 5.5 million downloads in the five days since the weights landed
The weights landed on August thirteenth. You hand the model images, video or words. The model writes words back. The licence lets anyone use it, including commercially. In the five days since, the four main copies of the model have been downloaded about five and a half million times. One community build went from under nine hundred thousand downloads to three and a half million in three days.
Answering a question and finishing a job on a machine are different skills
Plenty of models can answer a hard question. Very few can be handed a real computer and asked to finish a job on it. Rename these files. Fix this spreadsheet. Send that mail. Qwen3.8-27B is built for finishing the job on the machine. So how do you test whether a model can use a computer?
The model works a real desktop, running the same programs anyone runs
You sit the model in front of a real computer, a Linux desktop running the programs anyone runs: a browser, a spreadsheet, a mail client, a video player. The model gets one thing to work with, a picture of the screen. And the model answers with one thing, an action. Click here, at this exact point on the screen, in pixels. Drag this. Type that. So the model has to find the button as well as choose it. Every action carries a place on the screen. The machine carries that action out and hands back a fresh picture. That is one step. Then the model goes again.
At the end a script opens the machine and scores what it finds
The model gets a hundred of those steps at most. That is the setting the public scoreboard runs at, and a hundred steps is where the loop gets hard. One click lands wrong, and every picture after it shows a machine in a state the model never intended. The model has to notice that mistake and undo it, or it spends the rest of its steps acting on a plan the screen no longer matches. Then comes the grading. A checker program runs on the machine itself and reads its real state. Did the file get saved? Is the cell value right? The test is called OSWorld, and it ships a hundred and thirty-four of those checkers across three hundred and sixty-nine jobs.
One generation moved this model’s desktop score from 63.9% to 84.3%
One generation moved this model’s desktop score from 63.9% to 84.3%
These are Qwen's own numbers, measured on Qwen's own machines. On that run, Qwen3.8-27B averages eighty-four point three percent across those desktop jobs. The same team's model of the same size, one generation back, averaged sixty-three point nine percent on the same jobs. The gap between those two runs is twenty percentage points, at an identical model size.
Qwen has not entered 84.3 on the public scoreboard, where it would sit third
OSWorld keeps a public scoreboard that labs submit to, and a hundred and forty-four scores are on it. Qwen has not entered this score. If Qwen did enter it, eighty-four point three would come in third, below an agent system from Intelligence Indeed at ninety point two, and below Anthropic's Claude at just under eighty-six. One number does check out. The board lists Qwen's previous flagship at seventy-three point three, and Qwen's own card reports the same seventy-three point three.
The widest gap on its own card is a multidisciplinary exam set, 30.8 against 40.0
On every job that means operating something, a desktop, a phone, a browser, Qwen3.8-27B leads the models Qwen tested it against. On a multidisciplinary exam set, Qwen3.8-27B scores thirty point eight out of a hundred, against forty for the strongest of those baselines. Nine points down, the widest gap on its own card. Eighty-four point three is a desktop score. Thirty point eight is an exam score. Operating a computer is a skill of its own, and this card is the evidence.
Qwen’s desktop numbers were measured on weights anyone can download
Qwen’s desktop numbers were measured on weights anyone can download
The screen is the whole interface, and a point on it is the whole answer. The weights are fifty-five gigabytes, Apache licensed, and they sit on Hugging Face right now.
