H Company releases Holo4, a family of open-weight agentic computer-use models in 27B and 35B-A3B MoE sizes

H Company released Holo4, a family of open-weight "computer-use" agent models in two sizes (27B dense, 35B-A3B mixture-of-experts), built on top of Qwen base models, along with a smaller Holotron4 Nano variant. The models are downloadable in FP16, FP8 and GGUF formats and also available via H Company's API. Unusually, the company also open-sourced the step-by-step benchmark trajectories used to produce its performance claims, published at trajectories.hcompany.ai.

This gives developers and enterprises a self-hostable model that can drive software through GUI clicks, code execution, MCP and API calls without switching between separate specialised models, which removes real integration work. The bigger deal is the released trajectories: this is one of the few vendor benchmark claims that can actually be checked step by step rather than taken on faith. Impact is real but bounded, this is a cost and self-hosting play for teams building computer-use agents, not a capability leap that changes what's possible.

"Competitive with the frontier" oversells a roughly 20-point gap (61.7% vs 81.8% for Opus 5.5 on OSWorld 2.0), and the cost-performance charts use H's own harness and pricing with the frontier comparison line seemingly drawn to flatter Holo4. The weaker 35B-A3B model's 30.9% OSWorld score is easy to miss in the framing. This is also an incremental update on Holo3 riding Qwen's base model improvements, not a new capability class.

1. 01Independent reproductions of the OSWorld 2.0 and AutomationBench scores using the published trajectories, not H's own numbers
2. 02Actual adoption: Hugging Face download counts and any production deployments cited by third parties in the next 1-3 months
3. 03Whether competitors (Anthropic computer-use, OpenAI operator-style tools) respond with price cuts or open-weight releases of their own