Frontier Models
Agents-A1 Matches Trillion-Scale Performance with 35B Parameters via Horizon Scaling
InternScience at Shanghai Artificial Intelligence Laboratory open-sources a quantized 35B MoE model trained on extended agent trajectories to rival larger systems on benchmarks like SEAL-0 and IFBench.
Agents-A1 is a 35B Mixture-of-Experts agentic model developed by InternScience at the Shanghai Artificial Intelligence Laboratory that targets long-horizon agent tasks.
The release of Agents-A1 marks a notable shift in frontier model development toward optimizing for extended operational sequences rather than raw parameter counts. Traditional scaling laws have emphasized increasing model size to trillion parameters or beyond, yet this approach often incurs prohibitive compute and memory costs. Agents-A1 instead focuses on the quality and length of training trajectories to build capabilities in multi-step reasoning and tool use.
Shanghai Artificial Intelligence Laboratory positions the model as an accessible alternative for researchers and developers working on agentic systems. Built atop the Qwen3.5-35B-A3B base, Agents-A1 incorporates Mixture-of-Experts routing to maintain efficiency while handling complex workflows in search, engineering, and scientific domains. The open-source release under Apache 2.0 facilitates broad experimentation without licensing barriers.
What training methodology supports extended agent horizons in Agents-A1?
The development process centers on a knowledge-action infrastructure that generates lengthy agentic trajectories. These sequences average 45K tokens and span six heterogeneous domains, allowing the model to learn coherent behavior over prolonged interactions. This infrastructure supplies the raw material for subsequent fine-tuning stages that emphasize both breadth and depth of agent performance.
InternScience applied a three-stage recipe to transform the base model into an agent specialist. The initial phase conducts full-domain supervised fine-tuning to establish foundational instruction following across all target areas. Subsequent stages introduce domain-level teacher models that provide targeted guidance before a final distillation process routes examples through multiple teachers with vocabulary alignment to preserve salient tokens.
- Full-domain supervised fine-tuning across heterogeneous domains
- Domain-level teacher models for specialized guidance
- Multi-teacher domain-routed on-policy distillation with salient vocabulary alignment
This staged progression enables Agents-A1 to internalize patterns from long sequences without the need for commensurate increases in model size. The resulting system demonstrates improved coherence in tasks that require chaining multiple tool calls or maintaining context over extended research sessions.
How do benchmark results position Agents-A1 against larger models?
Performance evaluations highlight competitive outcomes relative to models with far greater parameter counts. On the SEAL-0 long-horizon search benchmark, Agents-A1 records 56.36. This figure surpasses or matches results from trillion-parameter systems referenced in the release materials. The IFBench instruction following benchmark yields 80.61, establishing state-of-the-art standing for the size class.
Comparisons drawn in the accompanying documentation place Agents-A1 ahead of Kimi-K2.6 and DeepSeek-V4-pro on these metrics despite the parameter disparity. Such outcomes suggest that trajectory length and training recipe can substitute for scale in agent-specific domains. Additional support for 256K context length further aids performance in scenarios demanding retention of earlier steps within a single session.
| Benchmark | Agents-A1 Score | Position vs Trillion-Parameter Models |
|---|---|---|
| SEAL-0 | 56.36 | Leading |
| IFBench | 80.61 | SOTA |
What deployment options does quantization provide for Agents-A1?
Following the initial June 26, 2026 open-source release, quantized variants appeared on July 2, 2026. These include GGUF formats that reduce memory footprint substantially. Lower VRAM requirements open access to consumer-grade GPUs and Apple Silicon devices, broadening the user base beyond institutional clusters.
Native integration of function calling and tool use remains intact across quantized checkpoints. This preservation allows developers to deploy the model in production agent loops without retraining or additional adapters. The Hugging Face repository provides direct examples for running on Mac hardware, underscoring the intent to democratize access to capable agent systems.
We introduce Agents-A1, a 35B Mixture-of-Experts Agentic Model that reaches trillion-parameter-level performance by scaling the agent horizon.Agents-A1 Team, Shanghai Artificial Intelligence Laboratory
What market and stakeholder implications arise from the Agents-A1 release?
The availability of a high-performing open agent model at modest size alters competitive dynamics in the frontier space. Organizations previously limited by hardware budgets can now experiment with long-horizon agents on local infrastructure. This accessibility may accelerate adoption in enterprise settings where data privacy precludes cloud reliance on larger proprietary systems.
Stakeholders in research and development gain a reproducible baseline for studying scaling laws specific to agent behavior. The Apache 2.0 license further encourages derivative work, including continued distillation or domain adaptation. Comparisons with GPT-5.5 and other closed models become feasible through direct evaluation rather than indirect inference.
What expert reactions and forward outlook surround Agents-A1?
The quoted statement from the development team emphasizes horizon scaling as a viable path to high capability. Documentation from the arXiv paper and repository reinforces that trajectory quality can deliver results previously associated only with extreme parameter counts. Community responses on release platforms have focused on practical deployment ease and benchmark gains.
Looking ahead, continued refinement of the training infrastructure could extend trajectory lengths further or incorporate additional domains. Integration with emerging agent frameworks may follow, alongside potential updates to the quantized line for even lower resource profiles. The release establishes a template for future models that prioritize efficient agent specialization over undifferentiated scale.
Frequently asked
How does the training of Agents-A1 differ from standard model scaling approaches?
Agents-A1 relies on long agentic trajectories averaging 45K tokens and a three-stage recipe of supervised fine-tuning, teacher models, and on-policy distillation rather than increasing parameter count alone.