Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent
arXiv · HuggingFace · ▲88
Abstract (verbatim)
We introduce Agents-A1, a 35B Mixture-of-Experts Agentic Model that reaches trillion-parameter-level performance by scaling the agent horizon. We investigate agent-horizon scaling from two perspectives: scaling long-horizon trajectories and scaling heterogeneous agent abilities. To support this goal, we build a long-horizon knowledge-action infrastructure that connects external knowledge, actions, observations, and verifier outcomes, producing agentic trajectories with an average length of 45K tokens. Based on this, we train Agents-A1 with a three-stage recipe. First, we perform full-domain supervised fine-tuning to align the base model with broad agentic behaviors. Second, we train domain-level teacher models to capture specialized expertise in each domain. Third, we propose a multi-teacher domain-routed on-policy distillation with salient vocabulary alignment to improve knowledge transfer efficiency across different domains, unifying six heterogeneous domains into one deployable student model. Agents-A1 achieves strong and broad performance for long-horizon agent benchmarks. Compared with 1T-parameter model such as Kimi-K2.6 and DeepSeek-V4-pro, Agents-A1 achieves leading results on SEAL-0 (56.4), IFBench (80.6), HiPhO (46.4), FrontierScience-Olympiad (79.0), and MolBench-Bind (56.8), and remains highly competitive on SciCode (44.3), HLE (47.6) and BrowseComp (75.5). We hope this work provides the community with a practical path for scaling the horizon using a 35B agent that can reach or match the performance of 1T models on long-horizon tasks.
Background
Background Analysis
1. Technical Context and Real-World Needs
Recent advances in large language models (LLMs) are shifting AI from passive language processing to autonomous agency, where models must plan, use tools, and adapt strategies in long-horizon tasks. These tasks—common in fields like software engineering, scientific research, and complex decision-making—require AI to operate over extended interactions, such as acquiring information, decomposing tasks, verifying intermediate results, and recovering from errors. The core need is to build systems that can solve problems persistently, not just respond to single queries. For example, a scientist might need an AI to design experiments, analyze data, and correct mistakes, while a developer might require an AI to write, debug, and optimize code iteratively.
2. Previous Limitations and Bottlenecks
Existing approaches to improve long-horizon agents follow two main paths:
- Parameter scaling: Models like Kimi-K2.6 or DeepSeek-V4 rely on trillions of parameters to internalize diverse reasoning and tool-use patterns. While effective, this requires enormous computational resources and is hard to replicate.
- Explicit horizon scaling: This approach aims to make decision-making processes explicit by training models on planning, tool interaction, and feedback. However, it faces two key challenges:
- Lack of knowledge infrastructure: Training long-horizon agents requires a unified environment connecting external knowledge, tools, observations, and verification signals. Without this, models struggle to learn from grounded feedback (e.g., verifying results or recovering from failures).
- Integrating heterogeneous abilities: Different domains (e.g., science vs. programming) demand distinct skills (e.g., information retrieval, constraint tracking), which are difficult to unify without conflicts.
3. Proposed Solution
The paper introduces Agents-A1, a 35B-parameter Mixture-of-Experts (MoE) model that scales agent capabilities by extending task horizons rather than parameters:
- Knowledge-Action Infrastructure: A unified environment connects external knowledge, tools, observations, and verification signals, generating long trajectories (average 45K tokens) for training. This allows models to learn planning, tool calling, and error recovery from real-world interactions.
- Three-Stage Training:
1. Full-domain supervised fine-tuning: Trains a general-purpose agent with broad long-horizon skills.
2. Domain-specific teacher models: Specializes models in individual domains (e.g., scientific computing).
3. Domain-routed on-policy distillation: Unifies skills from six domains into a single model using "salient vocabulary alignment" to avoid conflicts between reasoning patterns.
4. Key Differences from Prior Work
Unlike parameter-scaling approaches, Agents-A1’s innovations lie in:
- Solving infrastructure bottlenecks: Provides a dynamic environment for learning from feedback, not just text.
- Efficient integration of heterogeneous skills: Uses distillation to unify abilities (e.g., tool use, result verification) across domains, avoiding conflicts.
- Competitive performance with fewer parameters: Achieves trillion-parameter-level performance (e.g., outperforming Kimi-K2.6 and DeepSeek-V4 on SEAL-0 and IFBench) using only 35B parameters.
This work demonstrates a practical path to scaling long-horizon intelligence by optimizing task boundaries and skill integration, rather than relying solely on parameter growth.


