H Company announces Holo4 computer-use agents. The two-model family combines screen interaction, code and tool calls; its variants have different weight licenses, and H Company reports benchmark results with varying evaluation conditions. September 28, 2026 Key points * 01 H Company announced Holo4-27B, a dense model, and Holo4-35B-A3B, a mixture-of-experts model, on September 28, 2026. * 02 H Company says Holo4 can interact with graphical interfaces, run code and call Model Context Protocol tools and APIs. * 03 The downloadable weights have different licenses: CC BY-NC 4.0 for Holo4-27B and Apache 2.0 for Holo4-35B-A3B. * 04 H Company reports an 85.2% OSWorld score for Holo4-27B; 480 of AutomationBench's 600 public tasks were in the training-data split it used. H Company announced two Holo4 computer-use models on September 28, 2026, with distinct weight licenses, software tools and benchmark results. On September 28, 2026, H Company announced Holo4, a family of computer-use models with two variants: dense model Holo4-27B and mixture-of-experts model Holo4-35B-A3B. H Company says both are available through its H Models API and that their weights are published on Hugging Face. How Holo4 is designed to interact with software. H Company describes Holo4 as able to work through a graphical user interface (GUI), write and run code, and call tools through the Model Context Protocol (MCP) or APIs. The company lists desktop, web, Android, code-sandbox and business-API environments. That combination gives the models more than one way to act on a task: they can interact with screen controls, execute code or call connected tools, depending on the environment. These are capabilities H Company describes for the model family. Holo4-27B and Holo4-35B-A3B compared. | Variant | Model design | Parameters | Weight license | | Holo4-27B | Dense vision-language model, based on Qwen3.8 | 27 billion | CC BY-NC 4.0 | | Holo4-35B-A3B | Mixture-of-experts model, based on Qwen3.6 | 35 billion total; 3 billion active | Apache 2.0 | The licenses apply to the downloadable Holo4 weights. Holo4-27B's model card also identifies an Apache 2.0 license for its upstream Qwen3.8-27B model; that is separate from the CC BY-NC 4.0 license listed for Holo4-27B's weights. H Company also offers hosted access through the H Models API. What H Company's benchmark results mean. H Company reports an 85.2% OSWorld score for Holo4-27B and 80.8% for Holo4-35B-A3B. The company says its comparisons across models use different evaluation harnesses, effort settings and task subsets, so results across systems do not all come from the same conditions. Holo4's OSWorld 2.0 and ALE-CLI results use one run; most of its other benchmark scores average two to four runs. AutomationBench's public v1.0.6 set contains 600 tasks. H Company says 480 were in the split used to collect training data, and reports separate results on the remaining 120 held-out tasks: 49.3% for Holo4-27B and 31.7% for Holo4-35B-A3B. That distinction matters when assessing what the public-set figures say about performance beyond tasks in the training-data split. Training and the planned next release. H Company says supervised fine-tuning for Holo4 used 127 billion tokens, with about three quarters consisting of successful agent trajectories. It reports that the trajectories covered desktop tasks (45%), web (14%), MCP and API use (12%), and mobile tasks (3%). The company then trained two specialized LoRA experts and merged them into the fine-tuned model. H Company also says its Agentic Task Factory generated about 10,000 tasks: approximately 4,000 for web apps, 3,000 for MCP servers and 3,000 for desktop and operating-system tasks. The company planned to release optimized DSpark drafter checkpoints in the days following its September 28 announcement. Other Languages Read this article in another available language.

Holo4: powering generalist computer-use agents. Click, code and call tools across desktop, web and mobile. Holo4 27B scores 85.2% on OSWorld at $0.08 per task. Holo4 is its new series of agentic models. It comes in two sizes: 27B dense and 35B-A3B Mixture of Experts. Both are available on the H Models API. H Company is also releasing an updated version of Holotron 3: Holotron4 Nano. Holo4 builds on its previous model and interacts with software through any available interface: GUIs, code, MCP and APIs. It scores well on academic benchmarks, but H Company built it for real business workflows. It was trained through supervised and reinforcement learning on a large set of environments and tasks, including those generated by its Agentic Task Factory. Models built for every interface. Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools. It uses whichever fits the task. Most agentic models are trained for one interface only: GUI-focused models are blind without a screen, while models that prefer tool calling are stuck in front of an application that has no API. Real work is not siloed that way, and a single business task can require combining these different approaches. Holo4 runs on desktops, on the web, on Android, in a code sandbox and against business APIs. It is the same model in each case and it is called the same way. You do not need to select a different model for each platform. Holo4 models improve significantly over their Qwen base. 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%. However, it does so with orders of magnitude fewer parameters and at a much lower cost. H Company open-source every trajectory behind its scores on public benchmarks: replay each step at trajectories.hcompany.ai or download them from Hugging Face. Competitive with the frontier, at a fraction of the cost. On the hardest academic benchmarks for desktop control (OSWorld 2.0) and API use (AutomationBench), Holo4 competes with frontier models at a much lower cost per task. OSWorld 2.0 Average partial score (%) · higher is better * Holo * Base model * Reported * Closed frontier Lower cost USD per task · log scale AI that does work. Trained on environments and tasks from its Agentic Task Factory, Holo4 models excel on professional software. The examples below show Holo4 27B alongside Qwen3.8 27B, its base model. 84 calls · 1.3M tokens How H Company built Holo4. Agentic Task Factory. Its internal set of agentic pipelines builds interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. So far it has produced about 10,000 tasks across web apps, MCP servers and desktop environments, including hybrid environments that expose the same state through a GUI and MCP. Training. * 01 Supervised fine-tuning 127B tokens. About three quarters are successful agentic trajectories from its Agentic Task Factory: desktop (45%), web (14%), MCP and API (12%) and mobile (3%). The rest covers multimodal reasoning, GUI grounding, and text-only tool use and coding. * 02 Two RL experts Asynchronous online reinforcement learning on long-horizon tasks trains two specialized LoRA experts on the fine-tuned model: one for desktop and web, one for terminal, MCP and API. * 03 One merged model Both experts merge back into the fine-tuned model with equal weight and no further training. The result combines the general skills from supervised fine-tuning with each expert's specialized skills. Harness. Alongside training, H Company rebuilt its 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 their fixes. The largest changes were giving the agent a reliable memory that can keep track of hundreds of steps, and a shell on the desktop machine itself. Holotron4 Nano. Its post-training stack is designed to adapt to new foundation models and produce agents that generalize across interfaces and environments. As a member of the NVIDIA Nemotron Coalition, H Company applied its latest stack to the Nemotron 3 Nano Omni model as a follow-up to Holotron 3. The same recipe turns Nemotron 3 Nano Omni into Holotron4 Nano, a generalist agentic model that significantly improves over the base model on GUI workflows and in environments exposing MCP, APIs or coding sandboxes. These gains show that its recipe transfers well and can turn a generalist model into an agentic expert. Nothing in it is size-specific. Run it yourself. Both sizes are available today on the H Models API. Weights are on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF, next to its small model, Holotron4 Nano. H Company will release optimized DSpark drafter checkpoints in the coming days to further accelerate inference. Holo4 27B. Dense, 27B Best accuracy on long, multi-step tasks across web, desktop, and mobile. 85.2% on OSWorld * Context - 256K * Input / 1M$0.04 cached - $0.40 * Output / 1M - $3.00 Holo4 35B-A3B. MoE, 35B total, 3B active Close to 27B accuracy, cheaper and faster. Open weights under Apache 2.0. 80.8% on OSWorld * Context - 256K * Input / 1M$0.03 cached - $0.30 * Output / 1M - $2.00 Get started with Holo4. Open weights, a hosted API, and every run behind the numbers.

Laurent Sifre, cofounder of Paris-based agentic AI startup H Company, has joined Microsoft, Sifted has learnt. The move sees Sifre transition from the French AI startup to the US tech giant. H Company focuses on agentic AI technology. No further details about Sifre's role at Microsoft or the circumstances of his departure from H Company were disclosed.