Most scaling work adds parameters; this release shows you can instead scale the agent's horizon. Agents‑A1 was trained with long agentic trajectories (average ~45K tokens) and a multi-teacher, domain-routed distillation recipe so a 35B MoE can match many capabilities of much larger models while remaining deployable.
What Sets It Apart
- Long-horizon training target: trajectories produced by a knowledge-action infrastructure are orders of magnitude longer than typical fine-tuning data, enabling sustained multi-step planning and memory across tens of thousands of tokens. This is presented as the core lever for “trillion-parameter performance” without trillion-parameter weights.
- Architecture and packaging: a 35B Mixture-of-Experts base with quantized GGUF weights aimed at consumer-grade deployment. The model supports very long context lengths (serving examples use up to 262,144 tokens) and integrates native function/tool-calling workflows.
- Distillation & multi-domain teachers: three-stage training (full-domain SFT, domain teacher specialization, on-policy multi-teacher distillation with vocabulary alignment) unifies six heterogeneous domains into a single student model, improving cross-domain transfer while keeping model size manageable.
- Reproducible evaluation: published evaluation framework and benchmark numbers (e.g., strong results on IFBench, IFEval, FrontierScience tasks, MolBench-bind) accompany the release so users can reproduce comparisons under the same protocol.
Who It's For and Tradeoffs
Great fit if you want a locally-hostable agentic model that: requires long-context reasoning or multi-step tool workflows; needs compatibility with vLLM/SGLang; or you want a quantized GGUF build to run on commodity hardware. Look elsewhere if you need lowest-latency tiny models, strictly minimal memory footprints (MoE and long-context KV caches remain resource-heavy), or if your use case demands models trained specifically on a proprietary in-house dataset without further fine-tuning.
Where It Fits
This artifact is best treated as a deployment-ready, quantized model variant of an agentic research project: use it to prototype long-horizon agents, reproduce the paper's agent-evaluation, or integrate tool-enabled reasoning into products via vLLM/SGLang endpoints. It trades parameter count for trajectory-centered training and MoE sparsity to reach high benchmark performance while keeping model size practical.
