H Company launched Holo3-122B via API and Holo3-35B under Apache 2.0, with a claimed 78.9% OSWorld-Verified score and lower pricing than frontier peers. The release puts a new open computer-use model into direct benchmark competition with GPT-5.4 and Opus 4.6.
H Company shipped two variants. The model announcement describes Holo3-122B as its "most capable model" served through the API, while Holo3-35B is framed as a "lighter, faster variant" with "nearly the same intelligence" and an Apache 2.0 open release. For engineers, that makes this both a hosted frontier-model launch and an open-weights drop in the same family.
The headline metric is narrow but meaningful. H Company's launch post centers on a 78.9% OSWorld-Verified score, and the comparison chart in the benchmark summary lists Holo3-35B close behind at 77.8%. That same table shows Holo3 ahead of the cited Claude Sonnet 4.6 baseline on OSWorld-Verified, but not uniformly dominant elsewhere: WebArena is roughly tied with Sonnet 4.6, and the multi-app category favors Kimi-K2.5 and Claude Sonnet 4.6 over both Holo3 variants.
The cost story is more straightforward. H Company's pricing post puts Holo3-122B at $3.00/M output tokens and the open 35B variant at $1.80/M, while the launch graphic in the cost-performance chart positions both against much higher output-token prices for GPT-5.4 and Opus 4.6. If those benchmark conditions hold up independently, the interesting change is not just a new score leader on OSWorld-Verified, but an open 35B model entering the computer-use field with numbers close to its larger sibling.
Holo3 is here 🚀. Today, we're launching Holo3: our new series of frontier computer-use models. 78.9% on OSWorld-Verified. That puts us ahead of GPT-5.4 and Opus 4.6, at one-tenth of the cost. Weights on Hugging Face. API is live. Test it now! #Holo3 #OpenSource #ComputerUse Show more
