Holo4-35B-A3B
[!IMPORTANT] Holo4 family:
Model summary
Holo4-35B-A3B is a vision-language model (VLM) for Computer Use, built on Qwen3.6-35B-A3B and developed by H Company. Used with the hai-agents harness, it can send screenshots and tool results to the model, then execute its requested clicks, typing, code, and tool calls.
| Specification | Value |
|---|---|
| Model ID | Hcompany/Holo4-35B-A3B |
| Architecture | Qwen3.6 MoE |
| Checkpoint format | BF16 safetensors |
| Maximum context length in config | 262,144 tokens |
This demo shows Holo4-27B using FreeCAD to build a replica of the Eiffel Tower. More examples in the blog post.
PromptBuild a 3D Eiffel Tower in FreeCAD at a scale of 1 mm per metre. Center it on the origin and align it with the X and Y axes. Its plan must stay square at every height. The distance from the center to each corner is 62.5 mm at ground level, 32.5 mm at height 57, 17.5 mm at height 115, and 9.35 mm at height 276. Connect these widths with a smooth curve that narrows quickly near the base and more slowly near the top.
Make four separate, identical square legs, one in each quadrant. Their outer corners follow that curve, and each leg narrows from 14 mm across at ground level to 4 mm at height 276. Leave the space between the legs open. Add centered square platforms measuring 72 mm by 72 mm by 4 mm at height 57, 40 mm by 40 mm by 3 mm at height 115, and 22 mm by 22 mm by 3 mm at height 276. Add a square mast from height 276 to 324, tapering from 8 mm across to 2 mm across. Make every component a closed solid with nonzero volume, without filling the space between the legs.
Usage
The harness sends screenshots and tool results to Holo4, executes the model's requested actions, and sends the results back. It can give the model access to application tools and code execution.
Refer to the documentation for more details about:
Performance
Holo4 models improve significantly over their Qwen base models. On OSWorld, Holo4-27B scores 85.2% at $0.08 per task. We also evaluate them on Agentic Task Factory, a set of held-out business workflows across web, desktop, and MCP tools.
Benchmark results
Read the blog post for more tables about the benchmark results.
Pareto plots
These plots compare benchmark scores with cost per task. On OSWorld 2.0, Holo4-27B scores 61.7% at $1.22 per task, and Holo4-35B-A3B scores 30.9% at $0.61 per task.
On AutomationBench, Holo4-27B scores 45.4% at $0.05 per task, and Holo4-35B-A3B scores 34.5% at $0.02 per task.
Open-source evaluation traces
For transparency, we share all agent trajectories in the open-source dataset at Hcompany/trajectories.
Training
License
The model weights are available under the Apache License 2.0. This model is built on Qwen3.6-35B-A3B, which Alibaba Cloud also releases under the Apache License 2.0.