Holo3-35B-A3B
launchMarch 31, 2026
goblin vibe check:
open-source powerhouse if you need an ai to actually click buttons and test your game builds without paying per api call
state-of-the-art vision-language gui agent from h company tuned for autonomous browser, desktop, and mobile computer use with open-weight efficiency.
context
64k
tokens
cost
$0.25 in + $1.80 out per 1M
77.8% on OSWorld-VerifiedGUI agent specialized for browsers, desktops, and mobile interfacesMoE efficiency with 35B total parameters and only 3B active per stepStrong multi-step enterprise workflow execution across apps and tools
key features
77.8% on OSWorld-VerifiedGUI agent specialized for browsers, desktops, and mobile interfacesMoE efficiency with 35B total parameters and only 3B active per stepStrong multi-step enterprise workflow execution across apps and tools
spec & usage
Fine-tuned from the Qwen3.5-35B-A3B backbone for autonomous digital navigation
Runs interactively on a single 24GB consumer GPU such as an RTX 3090 or 4090
Native context length is 64,000 tokens and can extend toward 1M with scaling
Managed API pricing is about $0.25 per 1M input tokens and $1.80 per 1M output tokens
Weights are open under Apache 2.0 and the model is available through H Company's inference API
limitations
Specialized for computer use rather than broad general-chat workloads
The larger extended-context and enterprise workflow claims rely on the broader H Company stack, not just the base checkpoint alone
scope:
codesearchagentapilocalopen-sourcebenchmark-strongresearchfreemultimodal
launchMarch 31, 2026