Local AI is getting good, fast.
Open-weight models you can run on your own hardware are closing the gap on hosted AI month over month — and today's release is a good marker of how far that's come.
The gap is closing
A year ago, running something genuinely useful locally meant either a huge machine or a model that felt noticeably behind the hosted competition. That trade-off keeps shrinking. Each generation of open-weight models does more with less: smaller footprints, longer context windows, and reasoning that holds up on real, multi-step work instead of just benchmark trivia.
For a solo or small-firm practice, that matters. It's the difference between sending case material to someone else's server and keeping it on a machine you control.
Today's release: Qwen3.8-27B
Alibaba's Qwen team released Qwen3.8-27B today, open-weight under Apache 2.0. It's a dense, roughly 27.8-billion-parameter model with a 256K-token context window, and it's built to run on hardware people actually own rather than a rented server rack.
Quantized to 4-bit, it fits in around 17GB of VRAM — a single consumer GPU (24GB-class cards like an RTX 4090, or comparable Apple Silicon unified memory) is enough to run it. That's the kind of hardware a practice could plausibly already have, not a data center.
Example: one prompt, then a fancier pass
The best way to show what a model like this can do is to let it do something. Below is a small landing site for a fictional "Open Source Lawyer" project, built entirely on local hardware — no hosted API, no cloud inference. It went through two rounds.
Round one was a single prompt, unedited:
Hi, I want to make a website. The url is opensourcelawyer.com and the theme is using open source tools to help in
legal work. Two areas of focusing will be drafting with markdown and working with pdf files. Design the initial
landing page. If you have questions ask them. I want it to be as professional as possible. Save the website in this
workspace under a subdirectory of your choosing.
Round two asked for a fancier pass. The model preserved the original as a copy and asked which direction "fancier" should mean — visual richness, motion and interaction, live interactivity, or content depth. The reply:
ok. I'd like to make a much more fancy version but preserve this first shot. Can you copy the folder and make a new folder with "fancy" in the name first. Then we'll talk more.
[...]
premium consultancy, and increase the professionalism and impressive features in all areas 1, 2, 3, and 4.
What this means in practice
Privacy by default
Client material never has to leave the machine it's drafted on to get useful AI assistance.
No per-seat cost
Once the weights are on disk, there's no metered API bill for running them again.
Hardware you can budget for
A single consumer GPU is now enough for a genuinely capable model, not just a toy.