InternScience/Agents-A1-4B

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

InternScience/Agents-A1-4B is a 4.5 billion parameter agentic model developed by InternScience, designed to excel in long-horizon tasks across diverse domains. It is optimized for agentic reasoning, tool use, and instruction following, demonstrating strong performance in long-horizon search, engineering, scientific research, and general agentic tasks. This model achieves impressive results, often approaching or surpassing larger models, by scaling heterogeneous agentic abilities rather than just parameters.

Loading preview...

Overview

InternScience/Agents-A1-4B is a 4.5 billion parameter agentic model from InternScience, specifically engineered to handle complex, long-horizon tasks across various domains. This model is part of the Agents-A1 series, which focuses on scaling agentic abilities through a unique three-stage training paradigm and a domain-grounded knowledge-action infrastructure. It aims to achieve high performance with a significantly smaller parameter count compared to larger models.

Key Capabilities

  • Agentic Reasoning: Decomposes complex tasks, plans ahead, and adapts strategies based on intermediate results.
  • Tool Use: Natively supports function calling and integration with external tools like APIs, code interpreters, and search engines.
  • Scientific and Professional Reasoning: Handles tool-integrated scientific reasoning and professional knowledge-based question answering.
  • Instruction Following: Accurately follows detailed, multi-constraint instructions across diverse domains.

Performance Highlights

Despite its 4.5 billion parameters, Agents-A1-4B delivers strong performance across key agentic benchmarks. It significantly outperforms similarly-sized models on BrowseComp (66.8), XBench-DS-2510 (90.0), GAIA (95.1), FrontierScience-Research (33.3), and IFEval (94.8). Its scores on these benchmarks are competitive with, and in some cases surpass, larger MoE models like Nex-N2-mini and Qwen3.6, showcasing an excellent balance between efficiency and performance.

Good for

  • Developing local AI assistants requiring advanced agentic capabilities.
  • Applications needing robust long-horizon search and planning.
  • Tasks involving complex instruction following and tool integration.
  • Scientific research and engineering problem-solving where tool use is critical.

Finetunes
5 models