What happened
H Company released Holo4 on September 28, 2026, offering 27B dense and 35B-A3B MoE models with open weights on Hugging Face and API access. The models are designed for cross-platform computer use, supporting desktops, web, Android, and code sandboxes through a unified interface. The company reports the 27B model achieves 85.2% on the OSWorld benchmark at $0.08 per task, significantly outperforming its base model and competing with frontier closed models at lower cost. Weights are available under CC BY-NC 4.0 (27B) and Apache 2.0 (35B-A3B) licenses, with API pricing documented for both variants.
H Company released Holo4 on September 28, 2026, introducing two model sizes: a 27B dense model and a 35B-A3B Mixture of Experts (MoE) model. Both are available as open weights on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats, as well as through the H Models API. The 27B model is licensed under CC BY-NC 4.0 for research use, while the 35B-A3B model is under Apache 2.0, allowing for broader commercial application.
The models are built on the Qwen3.8 dense architecture and feature a maximum context length of 262,144 tokens. They are designed to interact with software through any available interface, including graphical user interfaces, code execution, MCP, and APIs. This unified approach allows the same model to operate across desktops, web browsers, Android devices, and code sandboxes without requiring platform-specific variants.
According to H Company, the Holo4-27B model scores 85.2% on the OSWorld benchmark at a cost of $0.08 per task, while the Holo4-35B-A3B scores 80.8% at $0.05 per task. These figures compare favorably to the base Qwen3.8 27B model, which scores 84.3% at $0.22 per task, and approach the performance of frontier closed models like Fable 5 (86.0%) and Qwen3.8 Max (86.1%) at significantly lower costs.
On the more challenging OSWorld 2.0 benchmark, which focuses on long computer workflows, Holo4-27B achieves a 61.7% score and 41.5% success rate at $1.22 per task. The company notes that while it trails the strongest closed models on long workflows, it does so with orders of magnitude fewer parameters and at a much lower cost. The company has open-sourced all trajectories behind its public benchmark scores for independent verification.
Why it matters
The release of Holo4 provides developers with a high-performance, open- alternative for building computer-use agents, reducing reliance on expensive proprietary APIs. By achieving near-frontier performance on OSWorld at a much lower cost per task, H Company demonstrates that smaller, specialized models can effectively handle complex agentic workflows. This shift could lower barriers to entry for enterprises and researchers looking to deploy autonomous agents for business automation, web navigation, and desktop tasks, while the open weights allow for local deployment and customization. The distinct licensing for the two model sizes also offers flexibility for commercial versus research applications.
The release of Holo4 is significant because it provides a high-performance, open- option for computer-use agents, a domain previously dominated by proprietary, expensive APIs. By achieving near-frontier performance on standard benchmarks like OSWorld at a fraction of the cost, H Company demonstrates that specialized, smaller models can be highly effective for agentic tasks.
The dual licensing strategy is particularly noteworthy. The Apache 2.0 license for the 35B-A3B model allows for unrestricted commercial use, making it attractive for enterprises looking to deploy agents in production environments without licensing fees. The CC BY-NC 4.0 license for the 27B model restricts commercial use, positioning it primarily for research and development.
The ability of Holo4 to operate across multiple platforms (desktop, web, mobile) through a single model reduces the complexity of deploying agentic systems. This cross-platform capability, combined with the open weights, enables developers to fine-tune and customize the models for specific business needs, potentially leading to more tailored and efficient automation solutions.
The transparency of the release, including the open-sourcing of benchmark trajectories and the detailed documentation of the training process, sets a high standard for reproducibility in the AI community. This transparency allows researchers and developers to verify the claimed performance and understand the underlying techniques, fostering trust and further innovation in the field of agentic AI.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').What most distinguishes an AI agent from a basic chatbot?
What to watch next
Independent verification of the reported benchmark scores, particularly on the held-out tasks of AutomationBench and the official private sets, is crucial to confirm the model's real-world efficacy. Developers should monitor the release of the promised DSpark drafter checkpoints for inference acceleration. Additionally, the practical utility of the model in long-horizon tasks, where success rates drop significantly compared to short-horizon benchmarks, will determine its viability for complex enterprise workflows. The adoption of the Apache 2.0 licensed 35B-A3B model in commercial products will be a key indicator of its market impact.
Independent verification of the benchmark scores is essential, as the company's results are based on its own harness and may not reflect performance in other environments. Researchers and developers should replicate the tests using the open-sourced trajectories and weights to confirm the reported figures.
The release of the DSpark drafter checkpoints, promised in the days following the announcement, will be important for assessing the model's inference speed and efficiency. Faster inference can significantly reduce the cost and latency of deploying agentic systems in real-time applications.
The practical performance of Holo4 in long-horizon tasks, where success rates are lower, will be a key factor in its adoption for complex enterprise workflows. Developers should monitor real-world deployments to see if the model can maintain reliability over extended periods of operation.
The adoption of the Apache 2.0 licensed 35B-A3B model in commercial products will be a strong indicator of its market impact. If major enterprises begin using Holo4 for their automation needs, it could signal a shift in the industry towards open- models for agentic tasks.