Beyond a Single Number: Evaluating Quantized Models for Deployment

I'm very interested to see how soon it will be possible to run Sonnet-class models (intelligence, not size) on relatively affordable computers such as my 32GB M1 Max Macbook. With the limited amount of memory available, I'm stuck with running quantized models (Q4 Qwen3.6 35B A3B, for example). What's frustrating is that there's very little objective information available about the performance hit caused by quantization. There's many subjective opinions to be found and so many variables that can influence the performance of the models. This article finally sheds a little bit of light on things. What I'm still missing, is a view on the effect of quantization on long-context tasks (>100k tokens), specifically agentic coding. I don't think the set of benchmarks used by ByteShape covers that.