It has indeed been really good, so much better than similarly sized models a few years ago. Would be interesting to know how far it is from what is possible to fit into a 27b model.

With which quantization ? I have a 16GB GPU, running Qwen3.5 9b UD_Q8_K_XL basically max out the VRAM usage. Maybe a MOE model fits you. Gemma 4 e4b is one with a total 8b.

And what is the token speed you are getting ?

I only have 8gb vram and can run 35b-a3b perfectly (30+ tps), have you tried 35b?

The setup described in this Codacus video is what got me started getting over 100k context with decent speeds on my RTX 3070 that I’ve been using daily for the last month: https://m.youtube.com/watch?v=0AqpaFm11oI.

The TheTom fork of llama.cpp adds asymmetric TurboQuant support that allows k at tq4 and v at tq2, which allows squeezing in more KV cache without quality loss. The REAP version of the MoE model also works just fine while further reducing the model size.

The TheTom fork doesn’t have pre-built container images, so it’s necessary to build your own (the ROCm Dockerfile in the devops directory).

If anyone does want to use the fork, I just recently built the ROCm image myself and ran into an issue where the fork’s Dockerfile was pulling a UI build package from Huggingface that doesn’t exist anymore, so I swapped in the latest ROCm Dockerfile from upstream that now builds the UI from source, and that worked fine.

How much CPU RAM do you have?

At 12GB you should be using a sparse MoE with offloaded experts. There are a few to choose from, but it just depends on how much spare RAM you have.

have you tried Qwen 3.6 35b a3b? check my guide, it's still relevant to you just with different numbers because you have 12GB

https://lemmus.org/post/24235317

Yeah do what these guys are suggesting and do the MOE with the experts offset. It really works well.

true, it's not perfectly clear

also I just saw this

OP's image is literally them saying something public about it