Shanghai AI Lab quietly put online, with no announcement, a browsing agent of around 744 billion parameters, freely downloadable. It would appear to beat GPT-5.6 Sol on BrowseComp, a test measuring the ability to find hard-to-locate information on the web. A public laboratory is releasing for free a model that beats the commercial product of a leading player on its own speciality.
We have been documenting the rise of open models for months. This episode adds an element we had not yet seen: specialisation.
What this test measures
BrowseComp evaluates one precise thing: the ability to find information on the web that is difficult to locate, by chaining searches, cross-checking sources and persevering when the first leads fail.
It is an agentic task in the strict sense, and it is different from pure reasoning. You have to decide what to search for, assess whether a result is relevant, rephrase when nothing comes up, and know when to stop. We described this chain in our article on web search.
Beating a frontier model on this exercise therefore does not mean being better in general. It means being better at that particular skill, which is already considerable.
What the silent release tells us
The manner of publication deserves as much attention as the result.
No announcement, no campaign, no comparison table put forward. The model is simply deposited, with its weights, and the community discovers it.
It is the exact opposite of what we covered this week: executives publicly debating the pace, announcements accompanied by in-house tables. Meanwhile, a laboratory deposits a 744-billion-parameter model without a word.
This echoes what we noted about the revelation of Ox Alpha: the brand matters less than one thinks, and professional developers adopt a model on its results before knowing where it comes from.
Until now, open models were nipping at the heels of closed models on generalist indices — a three-point gap, as we quantified in our comparison. Surpassing a frontier model on a precise skill is different. It suggests that a laboratory with more modest resources can target a segment rather than general performance, and come out ahead. This is exactly the logic of vertical specialisation that we described for companies, applied to research.
The usual caveats
One test does not make a model. BrowseComp measures one skill under given conditions. We recalled yesterday why benchmarks wear out: contamination, shortcuts, saturation. A high score on one precise test does not automatically carry over to real-world use.
744 billion parameters will not fit at home. This is the point we keep repeating: open does not mean accessible. A model of this size requires serious infrastructure, which reserves it for institutions or hosting providers.
No independent verification at this stage of the exact conditions of the evaluation.
What to take away
This episode illustrates a transformation we have been following since the summer and which is accelerating.
A year ago, the question was whether open models would catch up with closed models. It no longer arises in those terms: they are catching up on generalist indices and occasionally surpassing them in specialities.
What is at stake now lies elsewhere, and we wrote as much about model extraction: in a world where capabilities spread anyway, releasing voluntarily is not a commercial surrender. It is a way of capturing influence rather than trying to hold on to what is escaping.