H Company opens Holo4, a 27B model for every interface

H Company released Holo4, a family of agentic models built on a simple premise: rather than choosing between operating a graphical interface and calling tools, one model should be able to use whichever interface the task actually needs. Two sizes shipped, Holo4 27B dense and Holo4 35B-A3B, a mixture of experts model with 35B total and 3B active parameters. Both take a 256K context window and are available through the H Models API and on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF, with the 35B-A3B weights under Apache 2.0.

The third release is Holotron4 Nano, which is the same post-training stack applied to NVIDIA Nemotron 3 Nano Omni. That one is worth noting separately, because it is a signal that the training recipe matters more than the base model.

What the benchmarks show

• OSWorld: 85.2% for the 27B at $0.08 per task, 80.8% for the 35B-A3B at $0.05.
• OSWorld 2.0, the harder long workflow version: 61.7% for the 27B against 81.8% for Claude Opus 5.5, but at $1.22 per task instead of $8.48.
• AndroidWorld: 85.1% against 88.8% for Fable 5.
• AutomationBench: 45.4% at $0.05, and 49.3% on 120 held out tasks.
• ALE-CLI: 44.1% score, 19.4% pass rate.

Treat the 85.2% carefully. It is OSWorld, and H Company’s own reported harness. The OSWorld 2.0 figure of 61.7% is the more honest read on long horizon reliability, and the seven times cost difference against Opus 5.5 is the reason to care. API pricing is $0.40 per million input and $3.00 per million output for the 27B, $0.30 and $2.00 for the 35B-A3B, with cache reads at $0.04 and $0.03.

Training, and publishing the receipts

Training used the Agentic Task Factory, which generated roughly 10,000 verifiable tasks from documentation alone, split across about 4,000 web applications, 3,000 MCP servers and 3,000 desktop and OS environments. That was followed by 127B tokens of supervised fine tuning, then two asynchronous online reinforcement learning LoRA experts, one for desktop and web and one for terminal, MCP and API work, merged with equal weight.

The move that should earn trust is that every trajectory behind the public benchmark scores is being open sourced, at trajectories.hcompany.ai and on Hugging Face. H Company also flags that the public AutomationBench set includes tasks it trained on and gives the held out numbers separately, which is not something most launch posts do. Drafter checkpoints for speculative decoding were promised for the days after launch rather than shipped on the day.

(Original source: H Company)