trohrbaugh/Agents-A1-4B-heretic
The trohrbaugh/Agents-A1-4B-heretic model is a 4.5 billion parameter, decensored version of the InternScience/Agents-A1-4B model, created using Heretic v1.2.0+custom. This model is designed for agentic reasoning, excelling at decomposing complex tasks, planning, and adapting strategies. It features native tool use capabilities and strong performance in long-horizon search, engineering, scientific research, and instruction following tasks, often outperforming similarly-sized models and approaching larger MoE models in specific benchmarks.
Loading preview...
Model Overview
trohrbaugh/Agents-A1-4B-heretic is a 4.5 billion parameter language model, derived from the InternScience/Agents-A1-4B model and decensored using Heretic v1.2.0+custom. It is built for agentic reasoning, focusing on scaling heterogeneous agentic abilities across various domains. The original Agents-A1 series, developed by InternScience, emphasizes achieving high performance without relying on extremely large parameter counts, aiming for "trillion-parameter performance with a 35B agent."
Key Capabilities
- Agentic Reasoning: Decomposes complex tasks, plans, 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 & Professional Reasoning: Handles tool-integrated scientific reasoning and professional knowledge queries.
- Instruction Following: Precisely follows detailed, multi-constraint instructions across diverse domains.
- Decensored: This specific version has been modified to reduce refusals, showing 0/100 refusals compared to 99/100 in the original model.
Performance Highlights
Despite its compact 4B parameter size, Agents-A1-4B-heretic demonstrates strong performance across several benchmarks:
- Long-horizon Search: Achieves 66.8 on BrowseComp and 90.0 on XBench-DS-2510, outperforming Qwen3.5-4B.
- Engineering & Research: Scores 33.3 on FrontierScience-Research and 22.7 on MLE-Lite, showing competitive results.
- Instruction Following: Reaches 94.8 on IFEval, matching the flagship 35B Agents-A1 model.
Good for
- Developing Local AI Assistants: The 4B variant is designed to be faster and easier for building local AI assistants.
- Agent-based Applications: Ideal for tasks requiring complex task decomposition, planning, and tool interaction.
- Research & Development: Suitable for scientific reasoning, engineering tasks, and instruction-following applications where a smaller, efficient model with strong agentic capabilities is desired.