jbrahy/Qwen2.5-Coder-32B-Instruct-abliterated

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

jbrahy/Qwen2.5-Coder-32B-Instruct-abliterated is a 32.8 billion parameter Qwen2 dense architecture model, derived from Qwen/Qwen2.5-Coder-32B-Instruct, with a 32K context length. This version has undergone 'abliteration' using Heretic directional-refusal ablation to remove safety refusals, allowing it to generate content the base model would typically decline. It is primarily intended for use as a local coding assistant and for agentic coding applications, preserving the base model's strong coding abilities.

Loading preview...

Qwen2.5-Coder-32B-Instruct-abliterated: Uncensored Coding Assistant

This model is an abliterated (uncensored) version of the Qwen/Qwen2.5-Coder-32B-Instruct model, developed by jbrahy. It leverages the Qwen2 dense architecture with 32.8 billion parameters and a 32K context length, making it suitable for complex coding tasks.

Key Characteristics & Differentiators

  • Abliterated Safety Refusals: Created using Heretic directional-refusal ablation, this model has significantly reduced safety refusals, enabling it to respond to prompts the base model would typically decline. This process preserves the original model's coding capabilities without additional fine-tuning.
  • Strong Coding Ability: Inherits the robust code generation and understanding capabilities of its base model, making it effective for various programming challenges.
  • Flexible Formats: Available in bf16 safetensors for multi-GPU setups and various GGUF quantizations (Q6_K, Q5_K_M, Q4_K_M, Q3_K_M) to accommodate different hardware configurations, including GPUs and Macs.

Intended Use Cases

  • Local Coding Assistant: Ideal for developers seeking a powerful, less restrictive AI assistant for code generation, debugging, and understanding.
  • Agentic Coding: Well-suited for integration into agentic coding systems like opencode, where its uncensored nature can provide more direct and comprehensive responses.

Known Limitation

  • Tool Calling Format: While it reliably emits well-formed tool-call JSON, it does not wrap it in the <tool_call>…</tool_call> tags expected by OpenAI-compatible servers. This means tool calls may be surfaced as plain message content rather than dedicated tool_calls fields, requiring a thin proxy for agentic tool use.

Finetunes
1 models