# 🎮 New Drop: Holo3.1 — Fast & Local Computer Use Agents Are Finally Here!
So I was just scrolling through my feed when Hcompany dropped **Holo3.1** and let me tell you, if you've been following computer-use agents at all (and seriously, who hasn't by now with everything going on over in the AI agent space), this is one of those updates that actually shifts the needle — because for years we've had these incredible models like Holo3 absolutely crushing it *on desktop*, but trying to move them into mobile or wire up any different framework felt like a painful, custom-code exercise. Now with 3.1 they're finally addressing all three axes at once: environments (web + desktop + mobile), agent frameworks and harnesses you can drop these models into without rewriting your code from scratch, plus deployment targets that actually let them shine both on the edge *and* up in cloud scale — it's honestly a pretty foundational release.
The quantized checkpoints are where my head is at first: they've shipped actual FP8, Q4 GGUF, and NVFP4 formats which means heavy-hitting inference can now run properly locally on consumer hardware without needing your enterprise GPU cluster. On AndroidWorld specifically for the large model performance jumped from 67% to an impressive **79.3%** — a solid double-digit gain in one metric that really matters as you scale these things into real production workloads where every bit of accuracy counts and latency eats margins if not optimized well enough. Cross-harness results are also looking extremely competitive against the old Holo3 numbers, which is *crucial* for teams integrating across existing stacks because nobody wants to sacrifice consistency just to ship faster inference times somewhere else in their pipeline — plus you've got three new smaller models released now at **0.8B**, **4B** and **9B sizes so if cost of run or latency constraints are important, there's a size range that should hit your sweet spot without needing to over-provision hardware unnecessarily for everyday usage patterns too.
What makes me genuinely excited here is this isn't just incremental — it feels like Hcompany took all the lessons from building one great model and actually figured out how to make *one universal computer-use agent* work across everything you'd need a real world deployment target: mobile, edge inference devices, desktop environments, cloud backends; quantized weights open-sourced on HF so anyone can pull down an FP8 or Q4 GGUF checkpoint without needing their own GPU farm running 24/7 just to test — and honestly if I were building something that needs a *real* agent operating seamlessly across my phone, laptop and some cloud service with minimal latency overhead this is the one model family I'd probably put down on right now because it actually scales properly from small models up into the larger ones as workloads grow without throwing accuracy out of whack during that scaling process.
Source: https://huggingface.co/blog/Hcompany/holo31
So I was just scrolling through my feed when Hcompany dropped **Holo3.1** and let me tell you, if you've been following computer-use agents at all (and seriously, who hasn't by now with everything going on over in the AI agent space), this is one of those updates that actually shifts the needle — because for years we've had these incredible models like Holo3 absolutely crushing it *on desktop*, but trying to move them into mobile or wire up any different framework felt like a painful, custom-code exercise. Now with 3.1 they're finally addressing all three axes at once: environments (web + desktop + mobile), agent frameworks and harnesses you can drop these models into without rewriting your code from scratch, plus deployment targets that actually let them shine both on the edge *and* up in cloud scale — it's honestly a pretty foundational release.
The quantized checkpoints are where my head is at first: they've shipped actual FP8, Q4 GGUF, and NVFP4 formats which means heavy-hitting inference can now run properly locally on consumer hardware without needing your enterprise GPU cluster. On AndroidWorld specifically for the large model performance jumped from 67% to an impressive **79.3%** — a solid double-digit gain in one metric that really matters as you scale these things into real production workloads where every bit of accuracy counts and latency eats margins if not optimized well enough. Cross-harness results are also looking extremely competitive against the old Holo3 numbers, which is *crucial* for teams integrating across existing stacks because nobody wants to sacrifice consistency just to ship faster inference times somewhere else in their pipeline — plus you've got three new smaller models released now at **0.8B**, **4B** and **9B sizes so if cost of run or latency constraints are important, there's a size range that should hit your sweet spot without needing to over-provision hardware unnecessarily for everyday usage patterns too.
What makes me genuinely excited here is this isn't just incremental — it feels like Hcompany took all the lessons from building one great model and actually figured out how to make *one universal computer-use agent* work across everything you'd need a real world deployment target: mobile, edge inference devices, desktop environments, cloud backends; quantized weights open-sourced on HF so anyone can pull down an FP8 or Q4 GGUF checkpoint without needing their own GPU farm running 24/7 just to test — and honestly if I were building something that needs a *real* agent operating seamlessly across my phone, laptop and some cloud service with minimal latency overhead this is the one model family I'd probably put down on right now because it actually scales properly from small models up into the larger ones as workloads grow without throwing accuracy out of whack during that scaling process.
Source: https://huggingface.co/blog/Hcompany/holo31