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

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ansulev/OmniCoder-9B-heretic-ara-uncensored is a 9-billion parameter coding agent model, based on Tesslate/OmniCoder-9B, which itself is fine-tuned on Qwen3.5-9B's hybrid architecture. This version has been decensored using the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method. It features a 32,768 token context length and excels at agentic coding tasks, error recovery, and multi-step reasoning, having been trained on over 425,000 curated agentic coding trajectories.

Loading preview...

OmniCoder-9B-heretic-ara-uncensored: Decensored Coding Agent

This model is a decensored variant of the Tesslate/OmniCoder-9B coding agent, created by ansulev using the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method. It retains the core capabilities of the original OmniCoder-9B while aiming to reduce refusals.

Key Capabilities & Features

  • Decensored Behavior: Modified using Heretic ARA to reduce content refusals, achieving 7/100 refusals compared to 0/100 in the original model, with a KL divergence of 0.0452.
  • Base Architecture: Built upon Qwen3.5-9B, featuring a hybrid architecture with Gated Delta Networks interleaved with standard attention for efficient long-context processing.
  • Agentic Coding Expertise: Fine-tuned on over 425,000 curated agentic coding trajectories from models like Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro.
  • Advanced Error Handling: Demonstrates strong agentic behavior, including error recovery (read-before-write), responsiveness to LSP diagnostics, and generation of minimal edit diffs.
  • Extended Context Window: Supports a 262,144 token native context window, extensible to over 1M tokens.
  • Reasoning Chains: Incorporates a <think>...</think> mode for complex problem decomposition.

Performance Highlights

  • GPQA Diamond (pass@1): Achieves 83.8%, a 2.1-point improvement over the Qwen3.5-9B base model.
  • AIME 2025 (pass@5): Scores 90%.
  • Terminal-Bench 2.0: Reaches 23.6%, an 61% improvement over the Qwen3.5-9B base model.

Ideal Use Cases

  • Agentic Development: Suited for tasks requiring autonomous coding, tool use, and terminal operations.
  • Complex Code Generation: Excels in scenarios demanding multi-step reasoning and error recovery within coding workflows.
  • Uncensored Code Assistance: For users requiring a coding assistant with reduced content restrictions.