Holo4: powering generalist computer-use agents

What it claims

Holo4 is H Company's open-weight agent model pair: Holo4-27B (dense) and Holo4-35B-A3B (MoE). One model, called the same way, acts through GUI clicks and typing, self-written code, and MCP or API tool calls on desktop, web, Android, code sandboxes and business APIs. On OSWorld 2.0, Holo4-27B scores 61.7% against 81.8% for Opus 5.5. The headline pitch is cost per task, not absolute score. The MoE is much weaker on long desktop workflows at 30.9%, which the summary's framing of a "pair" hides. If you need OSWorld-class desktop control, the 27B is the one to evaluate.

How it works

• Agentic Task Factory. Internal pipelines build interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. The output is about 10,000 tasks across web apps, MCP servers and desktop environments. Some environments are hybrid, exposing the same state through both a GUI and MCP, which is how mixed-interface behaviour gets trained.
• SFT + RL on those tasks and others. The page gives no further training detail.
• Rebuilt harness. This is the loop that executes actions and manages context over hundreds of steps. It was tuned using OSWorld 2.0 failures: agents tagged why each task failed and engineers reviewed the fixes. The two largest changes were:
• a reliable memory spanning hundreds of steps
• a shell on the desktop machine itself

Scores are therefore partly a harness result. Expect lower numbers if you drop the model into your own loop.

Evidence

OSWorld 2.0 reference points

The authors state that releases, harnesses and task subsets differ across these points.

Cost comparisons

• Cost model: costs are estimated from tokens per run.
• Holo4 is priced at H Models API rates, from a single run.
• Qwen baselines are priced at Alibaba Cloud list prices.
• AutomationBench: no Holo4 numbers appear in the text. Holo4 was measured on v1.0.6 public tasks in H's internal harness, while competitors' costs come from the official leaderboard's private set. The two are not like-for-like.

Qualitative examples against the Qwen3.8 27B base (same prompt and harness):

The text gives no outcome or quality judgement for these runs, only call and token counts.

Holotron4 Nano reports gains as absolute percentage points over Nemotron 3 Nano Omni on GUI, MCP, API and coding tasks. The actual figures are not in the fetched text.

How to use it

• Hosted: both Holo4 sizes are on the H Models API. The page links a quickstart but gives no commands.
• Weights: on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF. The collection link says "FP16, FP8, GGUF", so check which precisions actually exist.
• Trajectories: every public-benchmark trajectory can be replayed at trajectories.hcompany.ai or downloaded as an HF dataset. These are useful for auditing scores or as demonstration data.
• Speculative decoding: optimized DSpark drafter checkpoints are promised "in the coming days".

Limits

• Licence: none is stated. Check the model cards before commercial use.
• AutomationBench: private-set results are pending, so the current comparison mixes public and private sets.
• OSWorld comparisons: these mix harnesses, subsets, effort settings and sources (model cards, the leaderboard, OpenAI's chart). Holo4 is excluded from the cost-frontier line.
• Single runs: Holo4's costs and scores come from a single run, with no variance reported.
• Missing details: the training section and the Holotron4 Nano gain table carry no numbers in the text.