Holo4 Agentic Models Launch on H Models API
Hcompany releases the Holo4 series—27B dense and 35B‑A3B Mixture of Experts—plus Holotron4 Nano, enabling agents that work across GUIs, code, MCP and APIs.
Original source published: September 28, 2026
Hcompany announced the Holo4 series of agentic models, offered in two sizes (27 B dense and 35 B‑A3B Mixture of Experts) and made available through the H Models API. An updated version of the earlier Holotron 3, called Holotron4 Nano, is also released alongside the main models.
The models interact with software via any interface—graphical user interfaces, code execution, MCP or API calls—and run unchanged on desktops, web browsers, Android devices, code sandboxes and business APIs. Trained with supervised and reinforcement learning on thousands of tasks from the internal Agentic Task Factory, Holo4 scores competitively on benchmarks such as OSWorld 2.0 (61.7% for the 27 B model) while operating at a fraction of the cost of larger closed models.
Weights are provided in BF16, FP8, NVFP4 and 4‑bit GGUF formats on Hugging Face, and the team will publish optimized DSpark drafter checkpoints soon. All trajectory data behind benchmark scores are openly released for reproducibility.