Thank you. This is an outstanding result.
In my personal tests, the model is on par with the iq4xs and may even be slightly better in programming. You’ve made a competent quant.
i tried same promt comparing this quant with Qwen3.8-27b-MoQ-GGUF 4.8, 2 attempts for both, Ridge failed 2 times, MoQ-4.8 succeded 2 times (kv cache F16, "medium" effort). Promt:
Write a single HTML file with a full-page canvas, no libraries. Animate a tree that grows from the bottom center of the screen in real time. The trunk grows upward first, then branches split off recursively with slight randomness in angle and length. Each generation of branches should be thinner and slightly lighter in color. When branches reach their final size, add small leaves as soft green circles at the tips. The tree should take about 15 seconds to fully grow. Use warm brown for wood and varied greens for leaves against a soft sky-blue gradient background.
This Ridge quant is much smaller, so it's expected... I'm not saying it's bad, i also tried different test promts i have, and it seems pretty smart.
I'm confused why you compared against the 4.8 instead of the 3.8 - 3.8 is the same size as this model and runs on 16gb. 4.8 is 40% larger and does not run on a 16gb card without offloading, which absolutely destroys performance on a dense model. The 4.8 MoQ almost certainly has the same or higher quants on the GDN that RIDGE has, so with higher quants everywhere else it sure as hell better beat RIDGE. Comparing against 4.8 just seems pointless, unless RIDGE somehow outperformed it (which again, it shouldn't).
I'm confused why you compared against the 4.8 instead of the 3.8 - 3.8 is the same size as this model and runs on 16gb. 4.8 is 40% larger and does not run on a 16gb card without offloading, which absolutely destroys performance on a dense model. The 4.8 MoQ almost certainly has the same or higher quants on the GDN that RIDGE has, so with higher quants everywhere else it sure as hell better beat RIDGE. Comparing against 4.8 just seems pointless, unless RIDGE somehow outperformed it (which again, it shouldn't).
I compared them because the models have different calibration. IQ4 is better at handling words and verbs. Ridge is a bit less accurate with simple word spellings and declensions, but in code and logic tasks, it delivers the same results; in one physics and programming task, it even performed better. On my 5060ti 16GB, I can easily run IQ4 at 25–30 t/s, but this model requires huge contexts, so Ridge is very good if you need it for coding.
empero-ai did an incredible job on this 27b ridge model, just wanted to say thank you as well.