Hcompany/Holo4-27B

TL;DR: Holo4-27B is H Company's open-weight (CC BY-NC) 27B dense computer-use VLM, fine-tuned from a Qwen base to drive an agent harness that loops screenshots and tool results in, and clicks, typing, code, and tool calls out, across web, desktop, and mobile. The headline claims are big — 85.2% on "OSWorld" at 0.08/task — but the card itself also reports 61.7% at1.22/task on "OSWorld 2.0", so the numbers need untangling before you believe the pitch. The genuinely notable move is releasing all evaluation trajectories as an open dataset, which lets you audit the claims yourself.

What it is

A vision-language model specialized for computer use — the "look at a screenshot, decide the next action" agent loop — shipped as one member of a three-model family (Holo4-27B dense, Holo4-35B-A3B MoE, Holotron4-30B-A3B on a NemotronH Nano Omni base). Each comes in BF16, FP8, NVFP4, and Q4 GGUF, so local deployment down to a single high-VRAM consumer GPU is plausibly on the table. Context is 262K tokens in the config, which matters for long multi-step trajectories full of screenshots.

The card says it's built on "Qwen3.8-27B" (Apache 2.0 base; the repo bundles both licenses). The model itself is CC BY-NC 4.0 — no commercial use, which limits adoption compared to Apache-licensed rivals like UI-TARS.

The numbers, and the catch

The card gives conflicting figures without explaining the difference:

85.2% at 0.08/task on OSWorld would be state-of-the-art territory at absurdly low cost. But the same card reports 61.7% at1.22/task on OSWorld 2.0. Most likely the 85.2% is the original OSWorld (now considered semi-saturated and known to have scoring bugs) and 61.7% is the harder 2.0 revision — but the card doesn't say whether these are verified runs, which harness config, or which subset. Also curious: the 35B-A3B MoE sibling scores dramatically worse (30.9% on OSWorld 2.0), suggesting the dense 27B is the model that got the real training attention.

The redeeming feature: all evaluation trajectories are public at Hcompany/trajectories. That's unusual and commendable — most computer-use leaderboard claims are unauditable. Before citing any number, pull a sample of trajectories and check the scoring, then reproduce a small OSWorld slice.

What's actually in the release

• Weights in four precisions (BF16 in this repo; FP8/NVFP4/GGUF as siblings in the Holo4 collection).
• Docs covering function calling, element localization, and document OCR — so it doubles as a grounding/OCR model, not just an end-to-end agent.
• The trajectory dataset and a blog post with fuller benchmark tables.
• Not in the card: any training details (the Training section is literally empty), the harness code location, or a runnable quickstart. That's a real gap — the model is only as good as hai-agents, and the card doesn't link installable code for it, just an API hub.

Verdict

Worth attention if you work on GUI agents: an open ~27B model claiming near-frontier OSWorld performance with quantized variants and auditable trajectories is exactly what the space needs. Treat the 85.2%/$0.08 figure as marketing until you've inspected the traces; the NC license blocks commercial use; and the empty training section plus harness dependency mean this is a "model drop" rather than a reproducible system.

Trying it

No snippet in the card. Shortest paths:

• Inspect the evidence first: trajectories.hcompany.ai / Hcompany/trajectories on HF.
• Harness/API docs: hub.hcompany.ai/models-api.
• Generic transformers load (standard image-text-to-text pattern; check the repo for the exact processor/chat template):

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "Hcompany/Holo4-27B", torch_dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("Hcompany/Holo4-27B")

For local use, the Q4 GGUF variant in the Holo4 collection is the cheapest entry point.