Scaling the horizon, not the model

For each task, InternScience trained on a full agent trajectory — every tool call, every observation it got back, every correction — averaging 45,000 tokens of decision-making rather than a question and a reply. The usual way to teach a model is to score its final answers; this scores the whole journey instead. They call it horizon scaling, and the pitch is that the axis worth scaling isn't the parameter count but the length and messiness of the problems the model is supervised through. The result, Agents-A1, is open under Apache 2.0.

The genuinely new move is treating the length of an agent's decision loop as the thing to scale — not the size of the network running it.

On the paper's own table it matches or beats three trillion-parameter frontier systems on a spread of long-horizon tasks — scientific olympiad problems, instruction-following, hard web search. That headline needs two asterisks the aggregator coverage dropped. The first: '35 billion parameters' is the memory footprint, not the working size. It is a mixture-of-experts model that lights up only about 3 billion parameters per step, so the real compute gap against a trillion-parameter incumbent is closer to a few hundredfold than thirtyfold. The second: the wins are lopsided. On machine-learning engineering it scores 43.9 where GPT-5.5 scores 72.7; on open web browsing it trails; it leads on the science and search benchmarks and loses on the tool-heavy ones.

The mechanism also leans on borrowing: the small agent is distilled from larger domain-teacher models, so part of the 'trillion-parameter performance' is trillion-scale knowledge compressed, not conjured from scratch. Read plainly, the claim isn't that a small model got smart on its own — it's that trajectory-shaped training plus distillation can pack frontier-grade agent behaviour into something you can self-host for roughly a tenth the token cost. For anyone running agentic pipelines on someone else's API, that trade — frontier-grade science and search behaviour at a tenth the token cost, in exchange for losing on tool-heavy work — is the whole decision, wins-and-losses and all; and because the weights and eval code are public, the split is something you can check yourself rather than take on the authors' word.