richardyoung/Agents-A1-4B-heretic

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The richardyoung/Agents-A1-4B-heretic is a 4.5 billion parameter language model with a 32768 token context length, created by richardyoung. It is a decensored version of InternScience's Agents-A1-4B, built using the Heretic v2.0.0.dev0 tool. This model specializes in agentic reasoning, tool use, and instruction following, with its primary differentiator being the removal of safety refusals through "abliteration" while retaining original capabilities. It is optimized for research, creative writing, and unrestricted local inference, particularly for tasks requiring long-horizon search, engineering, and scientific research.

Loading preview...

richardyoung/Agents-A1-4B-heretic: Decensored Agentic LLM

This model is a 4.5 billion parameter, 32768-token context length variant of InternScience's Agents-A1-4B, developed by richardyoung. It has been processed using the Heretic v2.0.0.dev0 tool to remove safety refusals, a technique known as "abliteration." This modification significantly reduces refusals (from 99/100 to 11/100) while maintaining the model's original capabilities, as indicated by a very low KL divergence of 0.060.

Key Capabilities

  • Decensored Output: Provides unrestricted responses by removing refusal mechanisms, making it suitable for research and creative applications.
  • Agentic Reasoning: Excels at decomposing complex tasks, planning, and adapting strategies for long-horizon trajectories.
  • Tool Use: Natively supports function calling and integration with external tools like APIs, code interpreters, and search engines.
  • Instruction Following: Precisely adheres to detailed, multi-constraint instructions across various domains.
  • Scientific & Professional Reasoning: Capable of tool-integrated scientific reasoning and professional knowledge question answering.
  • Performance: Despite its 4.5B size, it demonstrates strong performance on benchmarks like BrowseComp (66.8), XBench-DS-2510 (90.0), GAIA (95.1), and IFEval (94.8), often outperforming similarly-sized models and approaching larger MoE models.

Good For

  • Research: Ideal for exploring alignment techniques and conducting research without content restrictions.
  • Creative Writing: Supports creative applications requiring unconstrained text generation.
  • Unrestricted Local Inference: Suitable for local deployments where full control over model output is desired.
  • Agentic Workflows: Excellent for tasks involving multi-step reasoning, tool integration, and complex instruction following in domains like engineering and scientific research.

Finetunes
2 models