Hcompany/Holo4-35B-A3B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Holo4-35B-A3B is a 35.1 billion parameter vision-language model (VLM) developed by H Company, built on the Qwen3.6 MoE architecture. Optimized for computer use, it integrates with a harness to process screenshots and tool results, enabling automated execution of clicks, typing, code, and tool calls. This model excels at agentic tasks and business workflows across web, desktop, and multi-component tools.

Loading preview...

Model Overview

Holo4-35B-A3B is a vision-language model (VLM) developed by H Company, based on the Qwen3.6 MoE architecture. It is specifically designed for Computer Use, functioning as an agent that can interact with digital environments. The model leverages a specialized harness to interpret screenshots and tool outputs, subsequently executing requested actions such as clicks, typing, code execution, and tool calls.

Key Capabilities

  • Vision-Language Integration: Processes visual information (screenshots) alongside textual instructions.
  • Agentic Automation: Capable of automating complex business workflows across various interfaces including web, desktop, and other applications.
  • Tool Use: Can be integrated with application tools and code execution environments, allowing it to perform diverse tasks.
  • Performance: While Holo4-27B shows higher scores on OSWorld 2.0 (61.7%) and AutomationBench (45.4%), Holo4-35B-A3B offers a more cost-effective performance, scoring 30.9% on OSWorld 2.0 at $0.61 per task and 34.5% on AutomationBench at $0.02 per task.

When to Use This Model

  • Automating Digital Tasks: Ideal for scenarios requiring an AI to interact with a computer interface, such as data entry, software testing, or workflow automation.
  • Agent Development: Suitable for developers building AI agents that need to perceive and act within a graphical user interface.
  • Cost-Sensitive Agentic Workloads: Offers a balance of performance and cost-efficiency for agentic tasks, particularly where the 27B variant might be overkill or too expensive.

Training and Transparency

The model's training pipeline is detailed, and H Company provides open-source evaluation traces in the Hcompany/trajectories dataset for transparency. The model weights are released under the Apache License 2.0.

Finetunes
6 models