You can run local AI really fast
This blog post was released on the eighteenth of November 2026.
Strata engine
I decided to give Strata a try and it's absolutely zooming. Qwen 3.8 Flash Next with llama.cpp is faster than it was a few months ago but Strata is way faster. Furthermore, there is an RDNA3 fork of Strata and now I have 3k prompt processing speed and up to 90 tok/s for generation.
I knew that llama.cpp is general purpose and more optimized case-specific engines will come one day. It seems like this day has come and we will have only more different inference engines.
Some, of course, will run for speed and destroy the quality but it's obvious that inference can be better.
How's agentLIBRE?
agentLIBRE is doing really good, I use it every day and it works like magic. Of course, there are issues, problems, some bugs sometimes, sometimes something else to figure out but I have intelligence comparable to frontier models from six months ago, but on a high-end PC today.
One year ago local models were pretty dumb, while today they can power something like agentLIBRE and actually replace workflows from paid frontier models.
Qwen4 qwhen?
Tomorrow is a big day for local LLM enjoyers. Now I am an MoE believer, especially with Ngrams and alternative inference engines, but Qwen4 27B dense is still the extremely sweet release for everyone who runs local models.
I am of course excited to see the Flash model too, as 3.8 Flash Next is absolute magic for my config. At the same time, people with even 8GB VRAM GPUs or 32GB RAM can run beefy models. With more reasonably sized MoEs I think we will have fast inference on an average laptop. True acceleration.
Frontier models
I don't care about Astras, I tried Fable 5.2 when it was available in my plan, it was fine. Yes, it's smart but I don't feel like this smartness exceeds Opus 5.5 by a significant margin, so I'm not sad about not having Fable.
I use agl more and more, replacing one workflow after another. I also feel that the day when inference providers can actually provide value on a frontier AI lab level is closer and closer. For now, it's still an unexamined option for me.
Projects?
Well, I've started an everything-adjacent project and it's interesting and promising. I hope to show something as soon as possible, but for now agl receives the most attention.
REPL is better now, there is an untested-ish Windows version that works via WSL. Now agl is also nixified, I use the Nix language and package manager as much as possible.
There is much more but that's it for now. Focus is important and I'm happy to have interesting, and useful, projects to show. If you read this, this is a good opportunity to check agentLIBRE and maybe even give it a try.