H Company Releases Holo4: Open-Weight Computer-Use Models for Desktop, Web, Android and APIs

H Company announced Holo4, a family of computer-use vision-language models (27B dense and 35B-A3B MoE) built on Qwen bases; the 27B variant reportedly scores 85.2% on OSWorld at $0.08 per task.

H Company first published a Hugging Face blog post on 2026-09-28 titled "Holo4: powering generalist computer-use agents," but the article body was unavailable, leaving model details unconfirmed. A follow-up MarkTechPost report on 2026-09-29 supplied specifics: Holo4 is a family of computer-use vision-language models fine-tuned from Qwen bases, comprising Holo4 27B (dense, from Qwen3.8-27B) and Holo4 35B-A3B (mixture-of-experts with 3B active parameters, from Qwen3.6-35B-A3B), both with 256K context. A single model is said to handle GUI clicks, code, and MCP/API tool calls across desktop, web, Android and APIs. Reported benchmarks place Holo4 27B at 85.2% on OSWorld at $0.08 per task and 61.7% on OSWorld 2.0, versus Claude Opus 5.5's 81.8% at $8.48 per task (per MarkTechPost; per-report comparison fields are not directly reconcilable since the OSWorld 2.0 figure for Claude is given without a cost). Licensing differs by variant: the 35B-A3B weights ship under Apache 2.0 for commercial self-hosting, while the 27B weights are CC BY-NC 4.0. Training reportedly used a 10,000-task agentic task factory, 127B SFT tokens, and asynchronous online RL with two merged LoRA experts. All technical details come from the MarkTechPost report; the original Hugging Face post could not be verified because its body was unavailable.

• H Company announced Holo4 via a Hugging Face blog post on 2026-09-28; the post body was unavailable, so its contents could not be independently confirmed.
• Holo4 comprises two variants per MarkTechPost (2026-09-29): Holo4 27B (dense, fine-tuned from Qwen3.8-27B) and Holo4 35B-A3B (MoE with 3B active parameters, fine-tuned from Qwen3.6-35B-A3B); both have 256K context.
• Reported benchmarks: Holo4 27B scores 85.2% on OSWorld at $0.08 per task and 61.7% on OSWorld 2.0; Claude Opus 5.5 scores 81.8% at $8.48 per task.
• Licensing: Holo4 35B-A3B weights are Apache 2.0 (commercial self-hosting allowed); Holo4 27B weights are CC BY-NC 4.0.
• Training details per MarkTechPost: a 10,000-task agentic task factory, 127B SFT tokens, and asynchronous online RL with two merged LoRA experts.
• The models are described as single models handling GUI clicks, code, and MCP and API tool calls across desktop, web, Android and APIs.
• No source disagreements on facts; the only gap is that the primary Hugging Face source lacked a readable body.