H Company has unveiled Holo4, a groundbreaking series of AI models that can operate interfaces by clicking and typing, generate code, and invoke tools or APIs across desktop, web and Android environments. Holo4 is offered in two configurations—Holo4 27B (dense) and Holo4 35B-A3B (Mixture of Experts with 3B active parameters)—both supporting a 256K token context via the H Models API.
The larger 35B-A3B variant comes with Apache 2.0 open-source weights for full commercial self-hosting. The 27B model is released under CC BY-NC 4.0, restricting self-hosting for non-commercial use, with commercial access provided through H Company’s hosted API.
What Is Holo4?
Holo4 is a unified vision-language agent fine-tuned to interact with graphical and command-line interfaces, sandboxed code environments, MCP servers and RESTful APIs. Holo4 27B builds on Qwen3.8-27B, while Holo4 35B-A3B uses Qwen3.6-35B-A3B as its backbone. Both models integrate seamlessly with H Company’s open-source hai-agents Python harness, which feeds screenshots and prior tool outputs into the model and executes its click, type, code or API calls.
By combining GUI manipulation and direct API invocation in a single model, Holo4 addresses limitations of vision-only agents (which cannot act without a display) and tool-only agents (which stall when an API does not exist), delivering a consistent interface across desktop, web, mobile and server workflows.
Performance Benchmarks
According to H Company’s official benchmarks:
- OSWorld: Holo4 27B achieves 85.2% at $0.08 per task, compared to its Qwen3.8 27B base at 84.3% for $0.22.
- AndroidWorld: Holo4 27B scores 85.1% at $0.08 per task.
- AutomationBench (600 tasks): Holo4 27B reaches 45.4% at $0.05 per task. On the 120 held-out tasks unseen during training, it scores 49.3%.
For longer workflows on OSWorld 2.0, Holo4 27B records 61.7% at $1.22 per task, compared with Claude Opus 5.5 at 81.8% and $8.48 per task. H Company has published every evaluation trajectory at trajectories.hcompany.ai and on Hugging Face.
Training Pipeline
Agentic Task Factory
An internal pipeline generated over 10,000 verifiable tasks—comprising roughly 4,000 web app interactions, 3,000 MCP/server tasks, and 3,000 desktop and OS operations—by mining documentation and live software. Each task includes a verifier that rejects near-misses on purpose and requires success through the genuine interface.
Supervised Fine-Tuning
The supervised dataset totals 127 billion tokens, about 75% of which are successful agentic trajectories (45% desktop, 14% web, 12% MCP/API and 3% mobile), with the remainder covering multimodal reasoning, GUI grounding and text-only code/tool use.
Reinforcement Learning Experts
Two LoRA experts are trained asynchronously via online RL: one specialized in desktop and web, the other in terminal, MCP and API tasks. Both experts are then merged back into the fine-tuned checkpoint with equal weight to form a single unified model.
Enhanced Harness
The agent loop was overhauled using failure analysis from OSWorld 2.0: key improvements include reliable multi-step memory across hundreds of actions and an integrated shell environment for desktop operations.
Holotron4 Nano
H Company also introduced Holotron4 Nano, built on NVIDIA’s Nemotron 3 Nano Omni through the Nemotron Coalition. Holotron4 Nano elevates OSWorld performance from 21.0% in the base model to 76.3%.
Pricing and Deployment
- API Rates: Holo4 27B at $0.40 input / $3.00 output per million tokens; Holo4 35B-A3B at $0.30 input / $2.00 output.
- The API is OpenAI-compatible at
https://api.hcompany.ai/v1; see the quickstart guide. - Weights: Available on the Hugging Face collection in BF16, FP8, NVFP4 and 4-bit GGUF formats.
- Local Inference: Supported via vLLM or llama.cpp; DSpark drafter checkpoints for faster inference are forthcoming.
Interactive Explainer
Experience an in-depth walkthrough of Holo4’s agent loop, training recipe, benchmark economics and deployment options in our interactive explainer.
Feature Comparison
| Feature | Holo4 27B | Holo4 35B-A3B | Qwen3.8-27B | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|---|
| Architecture | 27B dense | 35B MoE (3B active) | 27B dense | Proprietary | Proprietary |
| License | CC BY-NC 4.0 | Apache 2.0 | Apache 2.0 | Closed | Closed |
| Context Window | 256K | 256K | 262K–1M | 1M | 1.05M |
| API Cost (1M tokens) | $0.40/$3.00 | $0.30/$2.00 | Varies | $4/$20 | $10/$50 |
| Interfaces | GUI, code, MCP, API | GUI, code, MCP, API | GUI, tools, code | Tools | Browser & tools |
| OSWorld 2.0 | 61.7% / $1.22 | 30.9% / $0.61 | 48.0% / $3.49 | 81.8% / $8.48 | 73.5% / $9.07 |
| Self-hostable | Non-commercial only | Yes | Yes | No | No |
*Cross-vendor scores use different harnesses; treat comparisons as directional.
Key Takeaways
- A single Holo4 model handles clicks, typing, coding and API/tool invocation.
- Holo4 27B reaches 85.2% on OSWorld at $0.08 per task; 61.7% on OSWorld 2.0 at $1.22.
- 35B-A3B offers Apache 2.0 weights for commercial self-hosting; 27B is CC BY-NC 4.0 (API only for commercial use).
- Holotron4 Nano boosts performance from 21.0% to 76.3% on OSWorld.
- API pricing starts at $0.30 input / $2.00 output per million tokens.
Learn more via H Company’s technical announcement, explore the Hugging Face collection or try the API at api.hcompany.ai.