Holo4 is H Company’s new series of agentic models, released in two sizes: a 27B dense model and a 35B-A3B Mixture-of-Experts model, both served through the H Models API. The release also includes Holotron4 Nano, a smaller model produced by applying the same post-training stack to the Nemotron 3 Nano Omni base.

The core design bet is interface generality. Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools, choosing whichever fits the task. The post argues most agentic models are trained for one interface only: GUI-focused models are blind without a screen, while tool-calling models are stuck in front of applications that have no API. Holo4 is the same model, called the same way, on desktops, the web, Android, code sandboxes, and business APIs.

Training used supervised and reinforcement learning on a large set of environments and tasks, including roughly 10,000 tasks generated by the Agentic Task Factory — internal agentic pipelines that build interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. These include hybrid environments that expose the same state through both a GUI and an MCP server. Alongside training, the team rebuilt the harness, the loop that executes the model’s actions and manages its context over hundreds of steps, using feedback from agentic performance on OSWorld 2.0. Agents tagged why each task failed and engineers reviewed fixes. The largest changes were giving the agent a reliable memory that can track hundreds of steps and a shell on the desktop machine itself.

On benchmarks, Holo4 improves significantly over its Qwen base but trails the strongest closed models on long workflows: on OSWorld 2.0, Holo4 27B scores 61.7% against 81.8% for Opus 5.5, and Holo4 35B-A3B reaches 30.9%, with orders of magnitude fewer parameters and much lower cost per task. Every trajectory behind the public benchmark scores is open-sourced, replayable via a viewer or downloadable from Hugging Face. The cost-performance discussion carries explicit caveats: costs are estimated from input and output tokens per agentic run, and releases, harnesses, and task subsets differ between reference points. On AutomationBench, scores and costs for Holo4 and the Qwen baselines were measured in the internal harness on v1.0.6, while other models use public-set scores from the README and the official leaderboard; a private-set evaluation of Holo4 is still pending.

Qualitative comparisons against the Qwen3.8 27B base, with the same prompt and harness, cover FreeCAD 3D modeling (an Eiffel Tower and an H company logo) and a self-playing Pac-Man-style game in Godot. The Eiffel Tower task shows Holo4 27B using 84 calls and 1.3M tokens versus 60 calls and 1.0M tokens for the base, while the Pac-Man task shows Holo4 27B at 68 calls, 2.4M tokens, and 268 lines versus 197 calls, 11.4M tokens, and 327 lines for the base.

Holotron4 Nano applies the same recipe to the Nemotron 3 Nano Omni model as part of the NVIDIA Nemotron Coalition, following up on Holotron 3. Reported gains are absolute percentage-point improvements over the base model on GUI workflows and in environments exposing MCP, APIs, or coding sandboxes, presented as evidence that the recipe transfers across foundation models and is not size-specific.

Both sizes are available today on the H Models API. Weights are on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF, alongside Holotron4 Nano. The post notes that optimized DSpark drafter checkpoints will be released in the coming days to further accelerate inference.