A 35B Model Matches Trillion-Parameter Performance by Scaling Horizon, Not Size
Agents-A1 reaches 1T-parameter-level benchmark scores by extending trajectory length and unifying heterogeneous tool domains, not by adding parameters.
The dominant assumption in agentic AI is that harder tasks require bigger models. More parameters, more capacity, better results. Agents-A1 breaks that assumption directly: a 35B Mixture-of-Experts model matches or beats trillion-parameter systems on long-horizon benchmarks by extending the agent's trajectory length and unifying its tool domains, not by scaling its weights.
The core mechanism is what the paper calls agent-horizon scaling, and it operates on two axes. First, trajectory length: instead of treating each task as a short, bounded interaction, the training infrastructure produces agentic trajectories averaging 45,000 tokens per sequence, connecting external knowledge retrieval, tool actions, environmental observations, and verifier feedback into a single coherent thread. Think of it less like a conversation and more like a work log, where the model accumulates evidence, corrects itself, and calls specialized tools across many sequential steps before committing to an answer. Second, heterogeneous capability unification: six distinct domains (including scientific reasoning, molecule binding prediction, and instruction following) are each handled by a specialist teacher model, then distilled into a single deployable student via multi-teacher domain-routed on-policy distillation with salient vocabulary alignment. That vocabulary alignment step is not cosmetic. Different domains produce different token distributions, and naive distillation loses signal at domain boundaries. Aligning the salient vocabulary across teachers before distillation preserves the transfer efficiency that makes the unified model competitive with each specialist individually.
Training follows three stages: broad supervised fine-tuning across all domains to establish agentic behavior baselines, domain-level teacher training to capture specialized depth, then the distillation stage that collapses six teachers into one model without requiring separate deployments per domain. The architecture stays at 35B active parameters throughout. No parameter count increase. The scaling happens entirely in the data infrastructure and the training recipe.
Agents-A1 scores 56.4 on SEAL-0, 80.6 on IFBench, 46.4 on HiPhO, 79.0 on FrontierScience-Olympiad, and 56.8 on MolBench-Bind, all leading results against 1T-parameter models including Kimi-K2.6 and DeepSeek-V4-pro. On SciCode (44.3), HLE (47.6), and BrowseComp (75.5), it remains competitive without leading. For teams evaluating whether to run trillion-parameter models for long-horizon agentic workloads, the takeaway is direct: a well-trained 35B model with extended trajectory infrastructure can match that performance at a fraction of the inference cost.
We're thinking: We read Agents-A1 as a direct empirical challenge to the inference-time scaling narrative that has framed most agentic research this year. The common framing puts inference-time compute on one side and training-time scale on the other, as if they are separate levers. What this work shows is that the two are entangled through trajectory design: if your training data contains long, tool-rich, verifier-grounded trajectories, the model learns to allocate inference-time compute productively across those steps. The 45K-token average trajectory length is not a side detail. It is the mechanism. Teams that invest in trajectory infrastructure, not just model size or inference-time search, may close capability gaps that currently seem to require orders-of-magnitude more parameters.
Key takeaways:
- Agents-A1 scales agent performance by extending trajectory length to 45K tokens average and distilling six heterogeneous domain specialists into one model via domain-routed on-policy distillation with vocabulary alignment, not by increasing parameter count.
- The 35B MoE model leads trillion-parameter baselines on five long-horizon benchmarks and remains competitive on three others, though the evaluation set is limited to tasks where long-horizon trajectory structure is the dominant performance driver.
- Teams building or procuring agentic systems for complex, multi-step workloads should audit their trajectory infrastructure before defaulting to larger models: the training data's trajectory length and tool diversity may matter more than the parameter count of the model being trained.