Agents-A1: A 35B Agent That Matches Trillion-Parameter Models by Scaling the Horizon

Bigger has been the reflexive answer to almost every question in AI. A new open-weights release from InternScience — a group within the AI for Science Center at Shanghai AI Laboratory — makes a pointed counterargument. Titled "Scaling the Horizon, Not the Parameters," the work introduces Agents-A1, a 35-billion-parameter Mixture-of-Experts model that reaches the performance of trillion-parameter systems on demanding agentic tasks. It landed on Hugging Face's daily papers on June 30 under an Apache-2.0 license.

Scaling the horizon, not the parameters

The central idea is "agent-horizon scaling." Rather than adding parameters, the team scales two other things: the length of the task trajectories a model learns from, and the breadth of specialized agent abilities it absorbs. To fuel this, they built a knowledge-action infrastructure that stitches together external knowledge, actions, observations, and verifier outcomes into coherent trajectories averaging 45,000 tokens — long, multi-step problem-solving episodes that resemble real agentic work far more than a typical question-answer pair.

A three-stage training recipe

Agents-A1 is trained in three stages. First, full-domain supervised fine-tuning aligns the base model with broad agentic behavior. Second, the team trains domain-level teacher models, each capturing deep expertise in one area — science, coding, browsing, and so on. Third, and most novel, is a multi-teacher domain-routed on-policy distillation step with "salient vocabulary alignment," which transfers each teacher's specialized skill back into the single student model efficiently. The result is a compact model that inherits a committee's worth of expertise.

Results and why it matters

Against 1T-parameter models such as Kimi-K2.6 and DeepSeek-V4-pro, Agents-A1 leads on several benchmarks — SEAL-0 (56.4), IFBench (80.6), the physics olympiad HiPhO (46.4), FrontierScience-Olympiad (79.0), and MolBench-Bind (56.8) — while staying competitive on SciCode, Humanity's Last Exam, and BrowseComp.

If the finding holds, it reshapes the economics of capable agents. A 35B model costs a fraction of a trillion-parameter one to run, so matching that tier through better training data and distillation — rather than raw scale — could put frontier-grade agents within reach of far smaller budgets. It is also another data point in a 2026 trend: China-origin labs releasing open weights that pressure the closed frontier from below.