H Company launches Holo3, claims 78.9% on OSWorld-Verified

• H Company launched Holo3, a computer-use model family aimed at GUI navigation and desktop agents, and the flagship 122B-A10B model claims 78.85 percent on OSWorld-Verified, according to the company launch materials launch post pricing summary.
• The more practical release for many teams is the open Holo3-35B-A3B model card, which is Apache 2.0 licensed, based on Qwen3.5-35B-A3B, and already tagged for Transformers support community summary model link.
• The benchmark chart matters because Holo3 is not just inching ahead, it is claiming a large jump over prior Holo and Qwen variants on OSWorld-Verified, with a much lower cost point than top proprietary models in the same chart cost-performance chart benchmark reaction.
• The fine print matters too: the strongest benchmark and pricing claims in circulation come from H Company materials and reposts, while the official docs are more concrete about API shape, multimodal inputs, and structured outputs than about independent evaluation context vendor claims model link.

The useful reveals are pretty simple: the official launch page says H trained Holo3 in synthetic enterprise environments, the quickstart docs show an OpenAI-compatible chat completions API with schema-constrained outputs, and the Hugging Face model card confirms the open 35B variant is a Qwen3.5 fine-tune under Apache 2.0. If you care about desktop agents, those three details matter more than the victory-lap tweets.

H Company built the launch around one figure: 78.85 percent on OSWorld-Verified for Holo3-122B-A10B. The official post positions that as state of the art on a benchmark for desktop computer use, and the comparison graphic places Holo3 above GPT-5.4, Sonnet 4.6, Opus 4.6, Qwen3.5, and Kimi in that specific setting.

That is the most important claim in the story, but it is still a vendor claim. If you want the benchmark context, BenchLM's OSWorld-Verified explainer is a decent refresher on what these tasks are actually testing: screen perception, stepwise action, state tracking, and recovery across real software interfaces.

The bigger practical move is the open Holo3-35B-A3B release. The model card describes it as a sparse MoE vision-language model for navigation and computer-use agents, with 35B total parameters and 3B active, fine-tuned from Qwen/Qwen3.5-35B-A3B and released under Apache 2.0.

That combination is rare enough to matter. Teams that want an inspectable GUI model now have a smaller open option instead of jumping straight to an API-only frontier model. The launch tweet also notes official Transformers support community summary, which lowers the friction for trying it in existing inference stacks.

H's quickstart and Models API docs show what the company thinks developers will actually do with Holo3:

• OpenAI-compatible Chat Completions API
• text plus image inputs
• structured text outputs, including schema-enforced JSON
• examples for coordinate prediction and UI localization
• a visible reasoning channel, with a docs switch to disable thinking output in extra_body

That last point is easy to miss. Holo3 is presented in the docs as a native reasoning model, and the examples explicitly show message.reasoning alongside normal content. For agent builders, that is more concrete than benchmark talk because it tells you how the model is expected to be wired into automation loops.

The launch page spends a lot of time on what H calls an "agentic learning flywheel." The ingredients are synthetic navigation data, out-of-domain augmentation, and curated reinforcement learning. H also says it built a proprietary "Synthetic Environment Factory" to generate enterprise-like websites and verifiable multi-step tasks.

That is the real story behind the benchmark jump. H is making a bet that desktop-agent progress comes from better synthetic task generation and verification, not just scaling a general model and hoping grounding improves on its own.

The comparison table in the shared launch material is a good reminder that Holo3 is strongest where H optimized it. The Holo3 variants lead on OSWorld-Verified, several single-app corporate categories, and ScreenSpot-Pro, but the same table shows weaker relative results on the multi-app category, where Kimi-K2.5 and Claude Sonnet 4.6 score higher community summary.

So the clean read is not "new best model, full stop." The cleaner read is that H seems to have built a very strong specialist for GUI navigation and grounded action, and the open 35B version makes that specialization accessible enough for engineers to test in real workflows.

H Company is pushing two tracks at once. The flagship Holo3-122B-A10B is available through H's inference API, while the 35B-A3B model is both in the API and openly released on Hugging Face under Apache 2.0 pricing summary model link.

That split makes sense. The 122B model is the benchmark spearhead, but the 35B model is the adoption wedge. If Holo3 gets traction outside demos, it will probably come from teams benchmarking the open 35B model on their own UI tasks, then deciding whether the flagship is worth paying for.