trohrbaugh/OmniCoder-9B-heretic-ara-uncensored

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

OmniCoder-9B-heretic-ara-uncensored is a 9-billion parameter decensored coding agent model, based on Tesslate/OmniCoder-9B and built using the Heretic v1.2.0 Arbitrary-Rank Ablation method. It is fine-tuned on 425,000+ agentic coding trajectories, leveraging Qwen3.5-9B's hybrid architecture with Gated Delta Networks for efficient long-context processing. This model excels at complex software engineering tasks, tool use, and multi-step reasoning, demonstrating strong error recovery and agentic behavior.

Loading preview...

Overview of OmniCoder-9B-heretic-ara-uncensored

This model is a decensored variant of Tesslate's OmniCoder-9B, created using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method. The original OmniCoder-9B is a 9-billion parameter coding agent built by Tesslate, fine-tuned on the Qwen3.5-9B base model, which features a hybrid architecture combining Gated Delta Networks with standard attention for efficient long-context processing.

Key Capabilities

  • Decensored Behavior: Modified to reduce refusals, showing 7/100 refusals compared to 0/100 in the original model, with a KL divergence of 0.0452.
  • Advanced Coding Agent: Fine-tuned on over 425,000 curated agentic coding trajectories derived from frontier models like Claude Opus 4.6, GPT-5.4, GPT-5.3-Codex, and Gemini 3.1 Pro.
  • Error Recovery & Agentic Behavior: Demonstrates robust error recovery (e.g., read-before-write patterns), responds to Language Server Protocol (LSP) diagnostics, and applies minimal edit diffs.
  • Long Context: Inherits Qwen3.5's 262,144 token native context window, extensible to over 1 million tokens.
  • Reasoning Chains: Supports <think>...</think> reasoning chains for complex problem decomposition.

Performance Highlights

OmniCoder-9B shows strong performance in coding and reasoning benchmarks:

  • GPQA Diamond: Achieves 83.8% pass@1 and 86.4% pass@3, outperforming the Qwen3.5-9B base model.
  • AIME 2025: Scores 90% pass@5.
  • Terminal-Bench 2.0: Achieves 23.6%, a significant improvement over its base model.

Good for

  • Software Development: Ideal for complex coding tasks, code generation, and debugging.
  • Agentic Workflows: Suitable for applications requiring autonomous agents that can recover from errors and perform multi-step reasoning.
  • Research & Experimentation: Useful for exploring decensored model behavior and the effects of ablation techniques on coding agents.

Finetunes
1 models