Mano‑P brings 4B GUI agent to Macs

GUI agents are the models that look at a screen, figure out what buttons and fields mean, and then click through software like a person would. The promise is huge(dev.to)he usual catch is ugly. Most of these systems send screenshots and task context to a cloud model. Mininglamp’s news is that it has now open-sourced a local Mac stack for this: Mano-P 1.0-4B, the Cider acceleration SDK, and Mano-AFK, all aimed at Apple Silicon. (dev.to) ### What actually shipped? Mininglamp didn’t just drop (dev.to)acceleration layer on top of Apple’s MLX stack, and Mano-AFK as an end-to-end automation builder. That matters because a lot of “open” agent launches really mean a checkpoint plus a demo. This one is trying to be a runnable stack. (dev.to) ### What is Mano-P? Mano-P is a GUI-aware vision-language-action model. In plain English, it reads what is on screen, reasons about the task, and outputs actions for desktop automation. Mininglamp says(dev.to) a staged training pipeline to make GUI interaction less brittle. The company is positioning the “P” as “Private,” which tells you the whole angle here. (dev.to) ### Why does running on a Mac matter? Because privacy is the whole sales pitch. If an agent is helping with finance dashboards, inter(dev.to)server is often the blocker. Mininglamp’s claim is that local mode keeps screenshots and task data on-device, with no cloud API required. That makes the Mac not just the client, but the inference box. (dev.to) ### What hardware does it need? This is not “runs on any laptop.” Mininglamp says direct deployment is currently for Apple M4 Macs with 32GB or more of u(dev.to) also points to testing on an Apple M5 Pro with 64GB RAM. So yes, it is local — but it still wants modern Apple Silicon and a decent memory budget. (dev.to) ### What does Cider do? Cider is the part that makes the local story more believable. MLX already supports some quantized inference, but Mininglamp says MLX lacks a true W8A8 path where (dev.to)mitives and Metal kernels. Basically, it is trying to squeeze more usable speed out of Mac hardware without punting the job to the cloud. (dev.to) ### Is the speedup a big deal? It is meaningful, but you should read it carefully. Mininglamp shows a roughly 12.7% prefill speedup for Mano-P 1.0-4B on an M5 Pro (dev.to)cond. Cider’s own README also shows operator-level gains that can be larger in some shapes. So the win is real, but this is more “make local inference practical” than “suddenly turns a Mac into a datacenter GPU.” (dev.to) ### How good is the model itself? Mininglamp highlights a 58.2% OSWorld success rate for the Mano-P 1.0 line an(dev.to)tached to the broader Mano-P 1.0 family, not clearly just the local 4B build. So the headline says the stack is competitive, but anyone evaluating it should separate family-level results from the smaller on-device model they can actually run. (dev.to) ### What’s the bottom line? The important shift is not just “another agent model launched.” It is that a company is packaging GUI aut(dev.to)to build on. The catch is hardware limits and early-stage tooling. But if local computer-use agents are going to become normal, this is what the stack starts to look like. (dev.to)