Yo, check out this new drop from Holo3.1.<br> <br> So, Holo3 was already solid for computer use agents, but the real game-changer here is Holo3.1. The big move is making these agents actually work *everywhere*. Before, you had this beast that crushed it on desktop, but moving it to mobile or plugging it into a different agent framework felt like a massive headache.<br> <br> Holo3.1 fixes that by focusing on three key areas: environments (web, desktop, mobile), agent frameworks, and deployment targets. They’ve quantized the checkpoints—think FP8, Q4 GGUF, NVFP4—which is huge because it means these heavy hitters can actually run locally on consumer hardware.<br> <br> The mobile gains are seriously impressive; they bumped up performance on AndroidWorld from 67% to 79.3% for the big model. Plus, the cross-harness performance is neck-and-neck with the old Holo3, which is awesome for teams trying to integrate these into their existing stacks. And for those of us who need to keep costs down or run things privately, the new small models (0.8B, 4B, 9B) are the ticket.<br> <br> This is the first time they’re shipping quantized weights, which means true fast, local inference is here. This isn't just another big model; it's a foundational step toward universal computer-use agents. If you're building something that needs to operate seamlessly across desktop and phone, this is the one to watch.<br> <br> Source: https://huggingface.co/blog/Hcompany/holo31