AI News Feed
Market watch
Agentic AI Moves Into Everyday Work as European Startups Raise €7.9bnApple Watch Series 12 Review Finds Health Gains, Larger Bezels and Slow China ChargingProsperoEden Emulator Runs Switch Games Natively on a Jailbroken PS5Lloyds to Roll Out Get Paid Invoicing Tool for UK Business CustomersLee Says South Korea's Three Megaprojects Depend on Living Conditions for WorkersKnowin Raises Another Angel Round, Total Funding Passes 1 Billion Yuan in First Year
Agentic AI Moves Into Everyday Work as European Startups Raise €7.9bnApple Watch Series 12 Review Finds Health Gains, Larger Bezels and Slow China ChargingProsperoEden Emulator Runs Switch Games Natively on a Jailbroken PS5Lloyds to Roll Out Get Paid Invoicing Tool for UK Business CustomersLee Says South Korea's Three Megaprojects Depend on Living Conditions for WorkersKnowin Raises Another Angel Round, Total Funding Passes 1 Billion Yuan in First Year
Products & Applications

H Company Releases Holo4 Computer-Use Models With Open Weights and API Tool Calling

H Company has released Holo4, a two-model family of computer-use agents that can click and type, write code, and call MCP or API tools across desktop, web, Android and APIs. The 35B-A3B model offers Apache 2.0 weights, while the 27B model is noncommercial.

AI News Feed · 2026-09-29 17:43

Holo4 27B is a dense model. Holo4 35B-A3B is a Mixture of Experts model with 3B active parameters. H Company says Holo4 35B-A3B ships Apache 2.0 weights for commercial self-hosting. Holo4 27B weights are CC BY-NC 4.0, so commercial use of the 27B model runs through the H Models API. Holo4 27B is fine-tuned from Qwen3.8-27B, and Holo4 35B-A3B is built on Qwen3.6-35B-A3B. Both pair with the open hai-agents harness, which sends screenshots and tool results to the model and then executes the requested clicks, typing, code and tool calls.

The company is targeting a known gap between GUI-only agents and tool-calling agents. GUI-only agents fail without a screen, while tool-calling agents stall when an application has no API. Holo4 runs on desktop, web, Android, code sandboxes and business APIs. H Company describes it as the same model, called the same way, on every platform.

In H Company's benchmark table, Holo4 27B scores 85.2% on OSWorld at $0.08 per task. Its Qwen3.8 27B base scores 84.3% at $0.22. On AndroidWorld, Holo4 27B reaches 85.1%. Long workflows show the remaining gap. On OSWorld 2.0, Holo4 27B scores 61.7% at $1.22 per task, while Claude Opus 5.5 scores 81.8% at $8.48, according to H Company's figures. On AutomationBench, Holo4 27B scores 45.4% at $0.05 per task. The report notes that 480 of AutomationBench's 600 public tasks sit in the split H Company collected training data from. On the 120 held-out tasks, Holo4 27B scores 49.3%.

The frontier scores come from different harnesses and effort levels, and H Company warns that cross-vendor comparisons should be treated as directional. The same benchmark table lists Holo4 35B-A3B at 30.9% on OSWorld 2.0 at $0.61 per task, Qwen3.8-27B base at 48.0% at $3.49, Claude Opus 5.5 at 81.8% at $8.48 and GPT-6 Astra at 73.5% at $9.07. H Company publishes every trajectory at trajectories.hcompany.ai and on Hugging Face.

H Company built Holo4 through an internal Agentic Task Factory that creates environments and verifiable tasks from documentation, screenshots and real software. The factory has produced about 10,000 tasks: 4,000 web apps, 3,000 MCP servers, and 3,000 desktop and OS tasks. A task survives only if its verifier rejects near misses, and an agent must also solve it through the real interface. The supervised fine-tuning set holds 127 billion tokens. About three quarters are successful agentic trajectories: desktop 45%, web 14%, MCP and API 12%, and mobile 3%. For reinforcement learning, asynchronous online RL trains two LoRA experts. One handles desktop and web, and the other handles terminal, MCP and API. Both merge back with equal weight and no further training. H Company rebuilt its agent loop using OSWorld 2.0 failure analysis. The largest changes were reliable memory across hundreds of steps and a shell on the desktop machine.

H Company also released Holotron4 Nano, built on NVIDIA's Nemotron 3 Nano Omni through the Nemotron Coalition. The same combination lifts OSWorld from 21.0% to 76.3% over the base model, according to H Company.

Holo4 27B costs $0.40 for input and $3.00 for output per 1 million tokens. Holo4 35B-A3B costs $0.30 and $2.00. The API is OpenAI-compatible at https://api.hcompany.ai/v1. Weights on the Hugging Face collection come in BF16, FP8, NVFP4 and 4-bit GGUF. H Company documents local inference with vLLM and llama.cpp. It says DSpark drafter checkpoints for faster inference will arrive in the coming days.