jbrahy/Qwen2.5-Coder-14B-Instruct-abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

jbrahy/Qwen2.5-Coder-14B-Instruct-abliterated is a 14.7 billion parameter Qwen2 dense architecture model, derived from Qwen/Qwen2.5-Coder-14B-Instruct. This model has undergone 'abliteration' via Heretic directional-refusal ablation, removing its safety refusals to generate content the base model would typically decline. Optimized for fast local use, it functions as an uncensored coding assistant, preserving coding ability while reducing refusal rates. It supports a 32K context length and is available in bf16 safetensors and GGUF Q4_K_M formats.

Loading preview...

Overview

jbrahy/Qwen2.5-Coder-14B-Instruct-abliterated is a 14.7 billion parameter model based on the Qwen2 dense architecture, specifically an 'abliterated' version of Qwen/Qwen2.5-Coder-14B-Instruct. Abliteration, performed using Heretic directional-refusal ablation, removes the model's safety refusals, allowing it to generate responses that the original base model would typically decline. This process aims to preserve the model's coding capabilities while significantly reducing its refusal rate.

Key Characteristics

  • Abliterated (Uncensored): Safety refusals have been removed, enabling it to answer prompts the base model would decline.
  • Base Model: Derived from Qwen/Qwen2.5-Coder-14B-Instruct, inheriting its core coding abilities.
  • Architecture: Qwen2 dense with 14.7 billion parameters and a 32K context length.
  • Formats: Available in bf16 safetensors (6 shards) and GGUF Q4_K_M (~8 GB), making it suitable for local deployment.

Intended Use & Limitations

This model is primarily intended as a fast local coding assistant. It is designed to fit comfortably on machines with around 18 GB of memory (e.g., Mac). A known limitation is its tool-calling behavior: while it emits well-formed JSON for tool calls, it does not wrap them in the <tool_call>…</tool_call> tags expected by OpenAI-compatible servers. This means tool calls appear as message content rather than a dedicated tool_calls field. For agentic use, a proxy is needed to convert the bare JSON. As a 14B model, it is also less reliable for long, multi-step agentic loops compared to larger alternatives.

Finetunes
1 models