darkc0de/Holo-3.1-35B-A3B-heretic

Overview

darkc0de/Holo-3.1-35B-A3B-heretic is a 35.1 billion parameter Vision-Language Model (VLM) that has been decensored from the original Hcompany/Holo-3.1-35B-A3B using Heretic v1.3.0. The original Holo3.1 family, developed by H Company and based on the Qwen 3.5 family, is specifically designed for computer use agents, offering capabilities for web, desktop, and mobile automation.

Key Capabilities & Features

• Decensored Version: Modified from the original Holo-3.1-35B-A3B to reduce refusals, with a reported 55/100 refusals compared to 97/100 for the original.
• Vision-Language Model: Integrates visual understanding with language processing for agentic tasks.
• Computer Use Agents: Optimized for automating interactions across various digital environments.
• Multi-Environment Support: Extends automation capabilities to web, desktop, and mobile platforms.
• Native Function-Calling: Supports seamless integration with agent frameworks.
• Optimized for Local Deployment: Available in various quantizations including BF16, FP8, NVFP4, and Q4 GGUF for efficient local inference.

Performance & Differentiation

The Holo3.1 family demonstrates strong performance across computer use, mobile automation, enterprise workflows, and UI grounding benchmarks. This specific 'heretic' version focuses on providing a less restrictive model while maintaining the core functionalities of the Holo3.1-35B-A3B, which is noted for its balance of performance and inference cost.

When to Use This Model

This model is suitable for developers and researchers requiring a powerful VLM for agentic applications, particularly those involving computer automation, UI interaction, and mobile control, where a decensored response profile is desired. Its local deployment optimizations make it a strong candidate for cost-efficient and privacy-sensitive applications.