H Company open-sources Holo4, computer-use agents for any UI
Two open sizes drive desktops, web, Android, code and APIs — Holo4-27B hits 61.7% on OSWorld 2.0 and runs on 79% fewer tokens than its Qwen baseline.
Copy markdown
Two open sizes, one interface layer
Holo4 ships as Holo4-27B (dense) and Holo4-35B-A3B (MoE, ~3B active per step), both open-weight on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF. One checkpoint drives GUIs, code, MCP and APIs across desktop, web, Android and sandboxes — enough to build a real computer-use agent this weekend without a frontier API bill.
79% fewer tokens, 65% fewer tool calls
In a Pac-Man build test, Holo4-27B finished in 2.4M tokens versus 11.4M for a Qwen baseline, with 65% fewer tool calls. In agent loops where tokens are the bill, that efficiency is the real draw — cheaper runs, not just open weights.
The benchmark reality check
On OSWorld 2.0, Holo4-27B scores 61.7% and the MoE variant 30.9%, against Opus 5.5's 81.8%. Frontier desktop control is still closed-model territory; Holo4's bet is competitive-enough accuracy at a fraction of the cost, with weights you can fine-tune on your own tasks.
Self-host or call the API, plus a Nano for the edge
Commercial use is permitted via H's Models API (pay-as-you-go) or full self-hosting, and H released evaluation trajectories with a step-by-step viewer for auditing. A lighter Holotron4 Nano, built on NVIDIA's Nemotron 3 Nano Omni, targets local and edge deployment.