Local throughput reaches 568 tokens per second using NVFP4
samsja#1768
Computer-use agents are moving from the cloud to your local machine. Fast.
When we launched Holo3 two months ago, the production feedback was clear: digital agents need to be blazing fast, cost-effective, and versatile.
Today, we're dropping Holo 3.1, engineered to run anywhere, instantly.
Massive token throughput. Low latency. Ready for your local workflow!
Many users are excited about Holo 3.1 because its local inference delivers strong computer-use performance on-device, sidestepping cloud latency, cost, and rate-limit problems.
No Digg Deeper questions have been answered for this story yet.
🌞This is big Local AI news! A new open-source Computer-Use LLM has just launched.
Holo 3.1 is H Company’s (🇫🇷) new local computer-use agent model that beats Qwen3.5-397B, Kimi-K2.5, and Sonnet 4.6!
Since it is built for local deployment → ⬩ Runs fully on your machine (MacBook, Windows PC, DGX Spark, RTX Spark) ⬩ Based on Qwen architecture, specialized for GUI understanding & computer control ⬩ Optimized checkpoints: NVFP4, FP8 & Q4 GGUF (0.8B to 35B sizes) ⬩ Strong gains: 79.3% on AndroidWorld benchmark (35B model)
💻 Comparison to Qwen3.5: Holo 3.1 is fine-tuned specifically for computer-use agents (screen understanding, planning, clicking, navigation). Better at real GUI tasks than general-purpose Qwen3.5, especially when running locally.⚡
Computer-use agents are moving from the cloud to your local machine. Fast.
When we launched Holo3 two months ago, the production feedback was clear: digital agents need to be blazing fast, cost-effective, and versatile.
Today, we're dropping Holo 3.1, engineered to run anywhere, instantly.
Massive token throughput. Low latency. Ready for your local workflow!
Holo 3.1 reaches a new SOTA on AndroidWorld, a popular computer use agents benchmark
Can be explored here https://paperswithcode.co/paper/85697
Computer-use agents are moving from the cloud to your local machine. Fast.
When we launched Holo3 two months ago, the production feedback was clear: digital agents need to be blazing fast, cost-effective, and versatile.
Today, we're dropping Holo 3.1, engineered to run anywhere, instantly.
Massive token throughput. Low latency. Ready for your local workflow!
📝 Technical Blog: https://hcompany.ai/holo3.1 🛠️ Holo Models API: https://hcompany.ai/holo-models-api 🤗 Hugging Face: https://huggingface.co/collections/Hcompany/holo31
@TeksEdge Basically they fine tuned Qwen3.5, why don't they compare themselves with Qwen3.6?
📦 On-Device Scale (0.8B to 35B) You no longer need massive cloud compute budgets. The Holo 3.1 family introduces ultra-lean 0.8B, 4B, and 9B configurations engineered specifically for low-cost edge deployment, scaling all the way to our flagship 35B model.
@TeksEdge 100 tokens/s even without mtp. On Dual sparks and doesn't get stuck on thinking like 3.6 Qwen does.
@TeksEdge Someone please throw this French shit on DeepSWE please
⚡ Fast & Local Inference (MacBook Pro to DGX Spark) We’ve shipped newly optimized NVFP4, FP8, and Q4 GGUF checkpoints designed specifically for high-performance local deployment. The speedups are substantial across the board: by partnering with NVIDIA to optimize inference speed, quantization on enterprise hardware like DGX Spark slashes average step times from 6.8s down to just 3.3s. Meanwhile, our new Q4 GGUF checkpoints bring those exact same workflows straight to consumer devices like a MacBook Pro.
📱 Breakthrough Mobile Performance Desktop was just the start. We’ve driven a massive leap in mobile environment control. On the AndroidWorld benchmark, our 35B model skyrocketed from 67% to 79.3%, while our compact 4B and 9B models leaped from 58% to 71%.
@hcompany_ai Couldn't find a 4bit 4B so I made my own GGUF on huggingface if anyone needs it.
@TeksEdge 🤔🤔
@TeksEdge beats them on what benchmark exactly. my local shell script also beats claude on 'not sending my tax pdf to amazon' but we don't press release that
@TeksEdge Hard pass for me on any cloud model with the keys to my digital life. Computer-use was always going to land local, but the model quality just wasn't there yet. Holo 3.1 closes that gap. This category's going local fast. Very exciting!
@hcompany_ai FUCKING HELL, YOU DID IT! <3
@TeksEdge It is built to be a good sub agent, comparing it to any of those models is misleading or worse, shows you are stupid.
@TeksEdge Local, French, runs on a MacBook. That's the actual headline I want to hear
@TeksEdge @mr_r0b0t u see this?
@Space_SIGINT @TeksEdge Yes very cool! Thinking about how this could be trained with Webwright traces 👀
@hcompany_ai And 0 thanks to qwen for the base models?
@TeksEdge yeah...why there's no qwen3.6 in that table? coincidance? I dont think so! 🤣
