coder3101/Qwen2.5-VL-72B-Instruct-heretic

VISIONPricing:Input $3.2 / Output $4Concurrent Unit Cost:4Model Size:72BQuant:FP8Context Size:32kPublished:Feb 21, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

coder3101/Qwen2.5-VL-72B-Instruct-heretic is a 72 billion parameter instruction-tuned vision-language model, a decensored version of Qwen/Qwen2.5-VL-72B-Instruct created using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA). This model excels at visual understanding, including analyzing texts, charts, and layouts within images, and is capable of processing long videos (over 1 hour) and performing visual localization with structured JSON outputs. Its primary use case is advanced multimodal reasoning, acting as a visual agent for tasks like computer and phone use, and generating structured outputs from visual data.

Loading preview...

Overview

coder3101/Qwen2.5-VL-72B-Instruct-heretic is a 72 billion parameter instruction-tuned vision-language (VL) model, derived from Qwen/Qwen2.5-VL-72B-Instruct. This version has been decensored using the Heretic v1.2.0 tool with Magnitude-Preserving Orthogonal Ablation (MPOA), resulting in a significant reduction in refusals (8/100 compared to 100/100 for the original model) while maintaining a low KL divergence of 0.0156.

Key Capabilities

  • Advanced Visual Understanding: Proficient in recognizing common objects and analyzing complex visual elements like texts, charts, icons, graphics, and layouts within images.
  • Agentic Behavior: Functions as a visual agent capable of reasoning and dynamically directing tools for computer and phone use.
  • Long Video Comprehension: Can understand videos exceeding 1 hour in duration, with the ability to pinpoint relevant segments and capture events.
  • Visual Localization: Accurately localizes objects in images by generating bounding boxes or points, providing stable JSON outputs for coordinates and attributes.
  • Structured Output Generation: Supports structured outputs for data from invoices, forms, and tables, beneficial for finance and commerce applications.
  • Dynamic Resolution and Frame Rate Training: Incorporates dynamic FPS sampling and mRoPE updates for enhanced temporal understanding in videos.

Good For

  • Applications requiring a less restrictive vision-language model.
  • Tasks involving detailed image analysis, including text and chart interpretation.
  • Developing visual agents for automated computer or phone interactions.
  • Processing and understanding long video content, identifying specific events.
  • Generating structured data from visual documents like invoices and forms.
  • Use cases demanding precise visual localization with programmatic output.

Finetunes
1 models