Impulse2000/OmniCoder-9B-heretic

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Impulse2000/OmniCoder-9B-heretic is a 9-billion parameter coding agent model, derived from Tesslate's OmniCoder-9B, which is fine-tuned on Qwen3.5-9B's hybrid architecture. This 'heretic' version has been decensored using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation, significantly reducing refusals from 82/100 to 1/100 while maintaining a low KL divergence of 0.0319. It excels at agentic coding tasks, error recovery, and multi-step reasoning, leveraging a 32768-token context window.

Loading preview...

OmniCoder-9B-heretic: Decensored Coding Agent

Impulse2000/OmniCoder-9B-heretic is a 9-billion parameter coding agent model, a decensored variant of Tesslate's OmniCoder-9B. It was created using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting and reducing model refusals.

Key Differentiators & Performance

  • Decensored Behavior: This 'heretic' version dramatically reduces refusals from 82/100 to just 1/100 compared to the original OmniCoder-9B, with a minimal KL divergence of 0.0319.
  • Agentic Coding Expertise: Inherits OmniCoder-9B's fine-tuning on over 425,000 curated agentic coding trajectories, including traces from advanced models like Claude Opus 4.6 and GPT-5.4.
  • Error Recovery: Demonstrates strong agentic behavior, including read-before-write patterns, responding to Language Server Protocol (LSP) diagnostics, and applying minimal edit diffs.
  • Hybrid Architecture: Built on Qwen3.5-9B's architecture, featuring Gated Delta Networks interleaved with standard attention for efficient long-context processing.
  • Context Length: Supports a native context window of 262,144 tokens (the base OmniCoder-9B supports 262K, and this model inherits that capability, though the prompt specifies 32768 tokens, the README indicates 262K).

Use Cases

  • Coding Agents: Ideal for applications requiring robust, agentic coding capabilities, including automated software development and debugging.
  • Complex Reasoning: Suitable for tasks that benefit from multi-step reasoning and error recovery in code generation.
  • Unrestricted Code Generation: Particularly useful for scenarios where a less restrictive model behavior regarding content generation is desired, given its decensored nature.

Finetunes
1 models