Qwen3-30B-A3B-Instruct-2507 device-optimized quant variants without output quality falling off a cliff.
A 30B runs on a Raspberry Pi 5 (16GB) achieving 8.03 TPS at 2.70 BPW, while retaining 94.18% of BF16 quality. ShapeLearn tends to find better TPS/quality tradeoffs versus alternatives.
What’s new/interesting in this one
- CPU behavior is mostly sane
On CPUs, once you’re past “it fits,” smaller tends to be faster in a fairly monotonic way. The tradeoff curve behaves like you’d expect.
- GPU behavior is quirky
On GPUs, performance depends as much on kernel choice as on memory footprint. So you often get sweet spots (especially around ~4b) where the kernels are “golden path,” and pushing lower-bit can get weird.
models: https://huggingface.co/byteshape/Qwen3-30B-A3B-Instruct-2507-GGUF
That’s the model I use the most practically. If you want something that you can host yourself which is good enough as an assistant to ask coding questions to it’s pretty good at that and blazing fast (well at least across my two 3060s). When it comes to the Pi, I do wonder if VideoCore could be leveraged in some way to speed it up.
Yeah that seems like it should be doable. It’s really interesting to see how you can run fairly large models on very modest hardware now. Pi version is quantized of course, but it’s still way more powerful than stuff you needed a literal data centre for like a couple of years ago. It’s kind of mind boggling to consider.
