H Company Releases Holo4 Open-Weight Agent Models

Holo4 Puts Clicks, Code and Tool Calls in One Model

Holo4 has arrived from H Company, and the release is a family of generalist computer-use models that treats the screen, the terminal and the API as one problem instead of three. The pitch is simple to state and hard to build: a single set of weights that clicks and types on a graphical interface, writes code, and calls MCP or API tools, without asking the operator to swap models when the job changes shape.

The family ships in two sizes. Holo4 27B is a dense model. Holo4 35B-A3B is a Mixture of Experts design with 3B active parameters. Both serve a 256K context on the H Models API. Deployability is part of the offer, but the two sizes do not come with the same license. The 35B-A3B ships Apache 2.0 weights for commercial self-hosting. The 27B weights are CC BY-NC 4.0, which means commercial use of that variant runs through the H Models API.

The Gap Holo4 Was Built to Close

Agent design has been split for a while, and the split has a cost. GUI-only agents fail the moment there is no screen to look at. Tool-calling agents stall when an application has no API to call. Holo4 is a vision-language model for computer use, and it is aimed squarely at that seam. The company targets a known gap: one model, called the same way, running on desktop, web, Android, code sandboxes and business APIs.

The 27B variant is fine-tuned from Qwen3.8-27B. The 35B-A3B is built on Qwen3.6-35B-A3B. Both pair with H's open hai-agents framework. The framework sends screenshots and tool results to the model, then executes the requested clicks, typing, code and tool calls. That loop is where a lot of agent projects quietly die, so the pairing matters as much as the weights.

Benchmarks: Close to the Frontier, Far From the Price

On H Company's benchmark table, Holo4 27B scores 85.2% on OSWorld at $0.08 per task.

Long workflows are where the remaining gap shows up. On OSWorld 2.0, Holo4 27B scores 61.7% at $1.22 per task. Claude Opus 5.5 scores 81.8% at $8.48, per H's figures. On AutomationBench, Holo4 27B scores 45.4% at $0.05 per task.

Frontier scores come from different harnesses and effort levels, so cross-vendor rows are directional at best. Opus 5.5 ran at max effort in Anthropic's setup. GPT-6 Astra ran at max effort on the 82-task offline subset. There is a second caveat that is easy to skim past: 480 of AutomationBench's 600 public tasks sit in the split H Company collected training data from. On the 120 held-out tasks, Holo4 27B scores 49.3%. H Company publishes every trajectory at trajectories.hcompany.ai and on Hugging Face.

How Holo4 Was Built

An Agentic Task Factory

Internal pipelines build environments and verifiable tasks from documentation, screenshots and real software. The factory has produced about 10,000 tasks: 4k web apps, 3k MCP servers, and 3k desktop and OS tasks. A task survives only if its verifier rejects near misses. An agent must also solve it through the real interface, not a shortcut around it.

Two RL Experts, One Merge

The supervised fine-tuning set holds 127B tokens. About three quarters are successful agentic trajectories, split across desktop at 45%, web at 14%, MCP and API at 12%, and mobile at 3%. Asynchronous online RL trains two LoRA experts. One handles desktop and web. The other handles terminal, MCP and API. Both merge back with equal weight and no further training.

The framework itself was rebuilt using OSWorld 2.0 failure analysis. The largest changes were reliable memory across hundreds of steps and a shell on the desktop machine.

Pricing, Weights and One More Model

Holo4 27B costs $0.40 input and $3.00 output per 1M tokens. Holo4 35B-A3B costs $0.30 and $2.00. The API is OpenAI-compatible at https://api.hcompany.ai/v1. Weights on the Hugging Face collection come in BF16, FP8, NVFP4 and 4-bit GGUF, and local inference is documented with vLLM and llama.cpp. H says DSpark drafter checkpoints for faster inference arrive in the coming days.

H Company also released Holotron4 Nano, built on NVIDIA's Nemotron 3 Nano Omni through the Nemotron Coalition. The same combination lifts OSWorld from 21.0% to 76.3% over the base model.

Holo4 uses one model for GUI clicks, code, MCP and API calls. That single-model claim is the whole argument, and the pricing is what makes it worth testing.

What to Watch

• Licensing splits the family: 35B-A3B is Apache 2.0, while 27B weights are non-commercial only.
• Holo4 27B hit 85.2% on OSWorld at $0.08 per task, but Opus 5.5 still leads OSWorld 2.0 at 81.8% versus 61.7%.
• API pricing starts at $0.30 input and $2.00 output per 1M tokens.
• The held-out AutomationBench number, 49.3% on 120 tasks, is the one to carry into any evaluation plan.

For teams already wiring agents into desktop software, the appeal is not a single benchmark row. It is the idea that Holo4 removes the handoff between looking at a screen and calling a tool. Whether that holds up on your workflows is a question the published trajectories are meant to help you answer before you spend anything.