H Company released Holo4 on September 28, 2026, introducing a series of agentic artificial intelligence models designed to operate software across graphical user interfaces, code sandboxes, APIs, and Model Context Protocol connections, according to a technical announcement published on Hugging Face.
The release provides two model architectures: a 27-billion-parameter dense model and a 35-billion-parameter Mixture of Experts configuration labeled 35B-A3B. Both systems handle tasks across desktop environments, web browsers, Android, and business APIs using identical model weights and call methods.
Developers built Holo4 on an Alibaba Qwen foundation, training the systems using supervised and reinforcement learning against roughly 10,000 interactive tasks produced by H Company's Agentic Task Factory. Engineers also redesigned the execution harness, equipping the models with direct desktop shell access and persistent memory to track actions over hundreds of sequential steps.
On the OSWorld 2.0 desktop benchmark, Holo4 27B scored 61.7 percent compared to 81.8 percent for Opus 5.5, while the Holo4 35B-A3B Mixture of Experts model reached 30.9 percent. In Godot game generation tests, Holo4 27B built an unattended Pac-Man implementation across 68 calls, 2.4 million tokens, and 268 lines of code, whereas Qwen3.8 27B required 197 calls and 11.4 million tokens.
Alongside the flagship release, H Company introduced Holotron4 Nano as an update to Holotron 3. The smaller model applies the company's post-training stack to NVIDIA's Nemotron 3 Nano Omni through the NVIDIA Nemotron Coalition.
H Company made the model weights available on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats alongside the H Models API. The company stated that optimized DSpark drafter checkpoints will launch in the coming days to speed up inference.
