Slide 1
A thirty-five-billion-parameter open model, with only about three billion active, just went toe to toe with trillion-parameter agents. It's called Agents-A1, from InternScience, and every weight is Apache-2.0.
They scaled the agent's horizon, not its parameters.
The trick isn't size. Instead of adding parameters, they scaled the agent's horizon: a loop of knowledge, an action, an observation, and a verifier that checks the result. Run it long enough and each training trajectory averages about forty-five thousand tokens.
Six expert teachers, distilled into one small student.
To make one small model broadly capable, they trained a separate expert teacher for each of six domains, from long-horizon search to scientific research. Then they distilled all six into one thirty-five-billion student, routing each rollout to the teacher that knows it best.
Slide 4
Here's the payoff, on the report's own numbers. On FrontierScience-Research it scores forty. GPT-5.5 scores twenty-six point seven. It also leads on instruction following and on Seal-0 search, and it lands GAIA at ninety-six, two points behind the best trillion-parameter model.
Slide 5
Now the honest part. Every baseline here is a number each lab reported for itself, not an independent re-run. And the trillion-parameter models still clearly win the hard stuff. On machine-learning engineering it's forty-four to seventy-three, and frontier coding and scientific tool use go the same way.
This week: a dense 4B you can run on a laptop.
And here's this week's twist. They shipped a dense four-billion version, with quantized builds that run on a laptop. It holds up: GAIA ninety-five to the big model's ninety-six, it ties on instruction following, and on a few tasks the little one wins outright. It only gives ground on the hardest search and engineering.
Slide 7
arXiv 2606.30616 · Apache-2.0
So the lesson isn't a bigger model. It's a longer horizon, and it put a laptop-size open agent into trillion-parameter territory. Weights and the full report are up now, Apache-2.0.
