H Company releases Holo4, a single model for clicks, code and API across platforms

H Company has released Holo4, an open-weight model family for computer use that can click, type, write code and call MCP tools or APIs from the same weight set. The two sizes — a 27B dense model and a 35B-A3B mixture-of-experts with 3B active parameters — both ship with a 256K-token context window on the company's API.

Holo4 is a vision-language model built on Qwen: the 27B variant is fine-tuned from Qwen3.8-27B, the 35B-A3B from Qwen3.6-35B-A3B. Both run on H's open hai-agents harness, which feeds the model screenshots and tool results and executes the actions it returns. The claimed advantage is a single model that operates on desktop, web, Android, code sandboxes and enterprise APIs without changing the call.

Benchmarks: near the frontier at a fraction of the cost. According to H's table, Holo4 27B scores 85.2% on OSWorld at 8 cents per task, versus 84.3% at 22 cents for the base model. On AndroidWorld it reaches 85.1%. The gap widens on longer horizons: on OSWorld 2.0 the model achieves 61.7% at $1.22 per task, while Claude Opus 5.5 scores 81.8% at $8.48 (per H data, under a different harness and maximum effort). On AutomationBench the score is 45.4% at 5 cents, though 480 of the 600 public tasks were included in the training data; on the 120 held-out tasks the score rises to 49.3%. All trajectories have been published for download.

How it was built: task factories and RL. The internal Agentic Task Factory generated roughly 10,000 verified tasks — 4,000 web, 3,000 MCP servers, 3,000 desktop and OS — each passing a verifier that rejects "near misses." The SFT dataset contains 127 billion tokens, approximately 75% successful agent trajectories (45% desktop, 14% web, 12% MCP/API, 3% mobile). In RL, two asynchronous LoRA experts were trained, one for desktop and web, another for terminal and API, then merged with equal weights and no further training. The harness was rewritten after OSWorld 2.0 failure analysis, emphasizing reliable memory over hundreds of steps and a shell on the desktop machine.

Holotron4 Nano and licensing. Also released is Holotron4 Nano, built on Nvidia's Nemotron 3 Nano Omni through the Nemotron Coalition; the same methodology lifts OSWorld from 21% to 76.3%. Licensing splits the family: 35B-A3B under Apache 2.0 (full commercial open-weight), 27B under CC BY-NC 4.0 (commercial use only via the API) — a notable distinction between open weights and open source.

Pricing and local deployment. On the OpenAI-compatible API: 27B costs $0.40 per million input tokens and $3.00 per million output; 35B-A3B costs $0.30 and $2.00 respectively. Weights are available on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF formats, with documentation for local inference on vLLM and llama.cpp. DSpark drafter checkpoints for faster inference are expected in the coming days.