Holo4 powers generalist agents capable of using a computer

Holo4 powers generalist agents capable of using a computer
3 min read

H presents Holo4, a new series of agentic models, meaning they are capable of acting autonomously on a computer or within software. It comes in two sizes, a 27B dense model and a 35B-A3B mixture-of-experts model, and both are available on the H Models API. The publisher is also releasing an updated version of Holotron 3, called Holotron4 Nano.

A single model for all interfaces

Holo4 builds on H’s previous model and interacts with software through all available interfaces: graphical user interfaces, code, the MCP protocol (which allows a model to connect to external tools), and APIs. It clicks and types on a screen, writes and executes its own code, or calls tools via MCP or APIs, depending on what suits the task. According to H, most agentic models are trained for only a single interface, with GUI-oriented models being helpless without a screen, and those favoring tool calls remaining stuck when faced with an application lacking APIs.

Holo4 works on computers, on the web, on Android, in a code sandbox, and with professional APIs. It is the same model in each case, called in the same way, so there is no need to switch models from one platform to another.

H trained Holo4 using supervised learning and reinforcement learning, across a large number of environments and tasks, some of which were generated by its Agentic Task Factory. This internal system builds interactive environments and verifiable tasks from documentation, for example screenshots of real websites or open-source software. It has so far produced around 10,000 tasks covering web applications, MCP servers, and desktop environments.

Announced performance against leading models

On OSWorld 2.0, a benchmark test measuring computer control, Holo4 27B scores 61.7%, compared to 81.8% for Opus 5.5. Holo4 35B-A3B reaches 30.9%. H believes Holo4 is only behind the most powerful closed models on long workflows, while using, according to the publisher, “orders of magnitude” fewer parameters and at a significantly lower cost.

On the most demanding tests for desktop control (OSWorld 2.0) and API usage (AutomationBench), H claims that Holo4 rivals leading models at a much lower cost per task. The publisher specifies that costs are estimated from the input and output tokens of each execution, and that the versions, environments, and task subsets differ from one ranking to another. H states it is publishing all the trajectories behind its scores on public benchmarks, viewable at trajectories.hcompany.ai or downloadable on Hugging Face.

In demonstrations compared to Qwen3.8 27B, its base model, Holo4 27B was evaluated on 3D modeling in FreeCAD, logo creation, and developing a Pac-Man-style game in Godot. For these examples, the publisher indicates the number of calls and tokens consumed by each model.

Holotron4 Nano and availability

H, a member of the NVIDIA Nemotron Coalition, applied its training method to the Nemotron 3 Nano Omni model, following Holotron 3. The result, Holotron4 Nano, is presented as a generalist agentic model, showing significant improvement on graphical workflows and in environments exposing MCP, APIs, or code sandboxes. The gains are expressed in absolute percentage points relative to Nemotron 3 Nano Omni.

Both sizes of Holo4 are available now via the H Models API. The weights are available on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats, alongside Holotron4 Nano. H also announces the upcoming release, in the next few days, of optimized DSpark “drafter” checkpoints, intended to accelerate inference.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x