Holo4: Open-Weight Agents Click, Code and Call APIs

H Company released Holo4, open-weight computer-use models that hit 85.2% on OSWorld at $0.08 per task while closing in on frontier prices.

@sharedot

Holo4: Open-Weight Agents Click, Code and Call APIs

By @sharedot · · 7 pages

  • AI Frontier
  • Open Weight Models
  • Computer Use Agents

H Company released Holo4, open-weight computer-use models that hit 85.2% on OSWorld at $0.08 per task while closing in on frontier prices.

What H Company shipped

H Company released Holo4, a family of generalist computer-use models for AI agents that click and type on screens, write code, and call MCP or API tools from one model.

It ships in two sizes: Holo4 27B, a dense model fine-tuned from Qwen3.8-27B, and Holo4 35B-A3B, a Mixture-of-Experts model with 3B active parameters built on Qwen3.6-35B-A3B.

Both serve a 256K context on the H Models API and pair with H's open hai-agents harness, which sends screenshots and tool results to the model and executes the requested actions.

MarkTechPost reports the target is a known gap: GUI-only agents fail without a screen, and tool-calling agents stall when an application has no API. Holo4 runs the same model the same way across desktop, web, Android, code sandboxes and business APIs.

Why the release is surprising

The surprising part is how close an open-weight model gets to the frontier at a fraction of the price.

According to MarkTechPost, H Company's benchmark table shows Holo4 27B scoring 85.2% on OSWorld at $0.08 per task, versus 84.3% at $0.22 for its Qwen3.8-27B base, and reaching 85.1% on AndroidWorld.

On longer OSWorld 2.0 workflows the gap remains: Holo4 27B scores 61.7% at $1.22 per task while Claude Opus 5.5 scores 81.8% at $8.48 per H's figures. MarkTechPost cautions that frontier scores come from different harnesses and effort levels, so cross-vendor rows should be treated as directional.

A secondary surprise: 480 of AutomationBench's 600 public tasks sit in the split H collected training data from, with Holo4 27B scoring 49.3% on the 120 held-out tasks.

How it was built

Per MarkTechPost, H Company built Holo4 through an Agentic Task Factory that produced about 10,000 verifiable tasks — 4k web apps, 3k MCP servers and 3k desktop and OS tasks — where a task survives only if its verifier rejects near misses.

The supervised fine-tuning set holds 127B tokens, roughly three quarters of them successful agentic trajectories. Training then used two asynchronous online RL LoRA experts, one for desktop and web and one for terminal, MCP and API, merged back with equal weight and no further training.

H also rebuilt its agent harness using OSWorld 2.0 failure analysis, with reliable memory across hundreds of steps and a shell on the desktop machine. The company also released Holotron4 Nano, built on NVIDIA's Nemotron 3 Nano Omni, which lifts OSWorld from 21.0% to 76.3% over the base model.

Stakes for open-weight builders

Holo4 35B-A3B ships Apache 2.0 weights for commercial self-hosting, while Holo4 27B weights are CC BY-NC 4.0, so commercial use of the 27B runs through the H Models API.

MarkTechPost notes pricing starts at $0.30 input and $2.00 output per 1M tokens for the 35B, and $0.40/$3.00 for the 27B, with an OpenAI-compatible API and weights on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF for local vLLM or llama.cpp inference.

This release also lands in a policy context: a Lawfare analysis argues that once open weights are published, export controls cannot recall the copies, though governments retain leverage over compute supply and downstream services like fine-tuning and hosted inference — a framework that shapes what releases like this can and cannot be restricted.

What comes next

H Company publishes every trajectory at trajectories.hcompany.ai and on Hugging Face, which MarkTechPost highlights as unusual transparency for a release this competitive.

The open-weight decision-model niche is also heating up in parallel: PR Newswire reports AutoTrust AI released JEV-27B, an Apache-2.0 model that answers structured decisions in a single forward pass on a frozen Qwen3.8-27B backbone, trained in about 9.2 hours on one NVIDIA B200.

Together the releases suggest builders will soon be able to self-host both the fast decisions and the screen-driving actions of an agent stack.

Sources

  1. marktechpost.com › H Company Releases Holo4: Open-Weight Computer-Use Models That Click, Code and Call Tools Across Desktop, Web, Android and APIs
  2. lawfaremedia.org › How Export Controls Can—and Cannot—Reduce the Risks of Open-Weight Models
  3. prnewswire.com › AutoTrust AI Releases JEV-27B, an Open Decision Model for Self-Hosted AI Agents

More on AI Frontier