Intel / article

Holo4: powering generalist computer-use agents

Source: huggingface.co
Date: Published Sep 28, 2026

From the Daily Brief

H Company released Holo4 in 27B dense and 35B-A3B mixture-of-experts variants. It is an open computer-use agent model that acts across graphical interfaces, code, MCP tools and APIs through a single interface, rather than specializing in screen clicking alone.

On OSWorld 2.0 it scores 61.7%, against 81.8% for Opus 5.5. That is a real gap, but Holo4 runs at a fraction of the parameters and cost, and H Company released full trajectories so users can inspect how it reaches its results rather than trusting a single number.

For teams building desktop or browser automation, Holo4 is a candidate for high-volume, lower-stakes steps where a frontier model is expensive, with escalation to a stronger model for the tasks it fails. The published trajectories also make it easier to judge failure modes before deployment.

Read the 2026-09-29 Brief →

Context

H Company designed Holo4 to combine screen interaction, code execution and software-tool calls within one model. This targets workflows that require different ways of operating applications: clicking and typing where a visual interface is needed, then using code or an API where programmatic access is available.

The company says it trained the models using supervised and reinforcement learning across varied environments. Its task-generation system produced about 10,000 tasks, including environments where a screen and tool interface expose the same underlying state. That provides training situations for moving between interfaces. The release makes model weights and action histories available, allowing developers to examine behavior beyond the headline scores.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

Read the source huggingface.co