H Company Ships Holo4 Agents, 27B Hits 61.7% on OSWorld 2.0
TL;DR
- Holo4-27B scores 85.2% on OSWorld in H's own harness and 61.7% on independent OSWorld 2.0, trailing Opus 5.5 at 81.8% -- a 20-point gap closed substantially at a fraction of the cost.
- AlphaSignal reports Holo4 uses 79% fewer tokens and 65% fewer tool calls than the base model, amplifying the per-task cost advantage beyond the headline $0.08 figure.
- The commercially licensed 35B-A3B Apache 2.0 variant scores 30.9% on OSWorld 2.0, roughly half the noncommercial 27B -- a genuine capability cliff for any enterprise team choosing self-hosting.
H Company's new Holo4 release ships in two sizes, a 27B dense model and a 35B-A3B mixture of experts, with weights on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF. On OSWorld 2.0, the 27B scores 61.7% and the MoE scores 30.9%. Anthropic's Opus 5.5 sits at 81.8% on the same board.
"Holo4 competes with frontier models at a much lower cost per task," the H team writes. The blog post charts average partial score against cost per task but lists no dollar figure in prose.
Alongside the models, H rebuilt its execution harness, "the loop that executes the model's actions and manages its context over hundreds of steps." The biggest changes were, in H's words, "a reliable memory that can keep track of hundreds of steps, and a shell on the desktop machine itself."
H, a member of the NVIDIA Nemotron Coalition, also applied the same recipe to Nvidia's Nemotron 3 Nano Omni to ship Holotron4 Nano, a follow-up to Holotron 3. Every trajectory behind the published scores is open-sourced at trajectories.hcompany.ai and as a dataset on Hugging Face.
What others are reporting
-
AlphaSignal Read →
Adds token efficiency metrics (79% fewer tokens, 65% fewer tool calls vs. base model) and a deployment matrix spanning managed API to edge quantization formats.
A single checkpoint can navigate menus in FreeCAD, call an enterprise API, and run a shell script within the same workflow.
-
Superpower Daily Read →
Leads with the licensing split as the core tension; uniquely flags AutomationBench data contamination concern (480 of 600 tasks from H's own training split) and questions cost comparison methodology.
The 27B model leads H's reported OSWorld 2.0 results at 61.7%, yet its weights are noncommercial.
-
DEV Community Read →
Covers training infrastructure depth: H's Agentic Task Factory generated ~10,000 verifiable hybrid-environment tasks; also notes persistent memory across hundreds of agentic steps.
Holo4 is trained to use one model across desktops, the web, Android, code sandboxes, and business APIs.
-
CC Leaks Read →
Emphasizes open evaluation trajectories at Hcompany/trajectories as an auditability differentiator vs. closed-model benchmarks; includes cost-performance ratio analysis across model sizes.
Holo4-27B scores 85.2% on OSWorld at $0.08 per task in H Company's harness.
Originally reported by huggingface.co
Read the original article →Original headline: H Company Open-Weights Holo4 Computer-Use Models, 27B Scores 85.2% on OSWorld at $0.08 per Task