Holo4 ships 27B and 35B-A3B agent models with open weights and trajectories

H Company has released Holo4, a pair of agentic models intended to operate across multiple interfaces instead of being trained around one narrow control channel. The family includes a 27B dense model and a 35B-A3B mixture-of-experts model. H Company says the same model can interact through graphical interfaces, execute code, call MCP tools and use APIs, which makes the release relevant to the growing overlap between open weights and agent infrastructure.

The source also publishes more than just an API endpoint. H Company links downloadable weights and says the collection includes BF16, FP8, NVFP4 and 4-bit GGUF variants. It also makes benchmark trajectories available for inspection. That is especially valuable for agentic claims because a single final score hides the sequence of actions, retries and tool choices that produced it. Public trajectories do not make a benchmark automatically comparable across harnesses, but they give researchers much more material to audit.

H Company reports strong results on computer-use and automation benchmarks and compares Holo4 with both open and closed systems. Open Model Weights should keep those results clearly attributed: harnesses, task releases and cost assumptions vary, and the source itself notes several of those differences. The durable facts for a registry are simpler and more useful: the weights exist, multiple formats are published, the models target cross-interface agent workflows, and the release includes unusually inspectable execution traces.