Agents A1: A 35B agent model punching above its weight

Agents-A1 is an open-source 35B mixture-of-experts model from InternScience, engineered specifically for multi-step autonomous workflows and heavy tool calling.

Unlike traditional conversational LLMs, it is designed from the ground up as an agent. You give it an objective, and it autonomously breaks down the problem, triggers external search engines or code interpreters, observes the intermediate output, and course-corrects until the job is done.

What it actually does

InternScience built Agents-A1 to scale long-horizon execution rather than raw parameter count. The team used a three-stage training recipe: broad agent fine-tuning, domain-specific teacher models, and multi-domain on-policy distillation with action-verification feedback.

The benchmark numbers are striking for a ~35B footprint. On the GAIA benchmark for general agentic tasks, it scored 96.0. It hit 94.8 on IFEval and 80.6 on IFBench for instruction following, alongside 75.5 on BrowseComp and 86.0 on XBench-DS-2510.

Its scientific reasoning is where it punches furthest above its weight. With tool use enabled, it scored 47.6 on Humanity’s Last Exam, 79.0 on FrontierScience-Olympiad, and 46.4 on the HiPhO physics benchmark. InternScience reports that it outscores Qwen 35B across the board and edges past trillion-parameter giants like GPT-5.5, Kimi-K2.6, and DeepSeek-V4-pro on several specialized benchmarks.

How to run it locally

Weights, evaluation code, and technical reports are available on Hugging Face under InternScience/Agents-A1, with full compatibility across Transformers, vLLM, and SGLang. It supports up to a 262K context window.

Serving it locally is straightforward. With SGLang, a single CLI command spins up an OpenAI-compatible endpoint with built-in reasoning and tool-call parsers.

Hardware requirements are flexible. The unquantized model takes around 70GB of storage. If you are constrained on VRAM, there is an official FP8 release at 37.7GB, and a Q4 quantized version sitting at just 21GB that easily fits on high-end consumer GPUs. Quantized builds are also available via the MLX community for Mac users, and a smaller 4B model has been released for even lighter setups.

Why it matters

If you are building autonomous coding, research, or workflow agents and want to avoid massive cloud API bills or sensitive data leaks, Agents-A1 is worth spinning up immediately. Running true agentic planning on a single local GPU used to be a compromise, but this model closes that gap significantly.

We think this release signals an important shift in open-source AI: brute-force scale is no longer the only way to build great reasoning agents. Deep distillation focused on tool-use trajectories allows a 35B model to challenge trillion-parameter frontier systems in domain-specific tasks.