Controlled comparison
Q4_K_M vs Q8_0: does the bigger quant buy anything?
Is it worth spending the extra memory on the Q8_0 artifact of Qwen3.6-35B-A3B for chat and JSON automation on Strix Halo?
What was held fixed
Same unit, same Vulkan backend build, same frozen corpora, single stream, reasoning off. Only the model artifact differs (both pinned by hash).
| Metric | Q4_K_M (20 GB) | Q8_0 (35 GB) |
|---|---|---|
| Inter-token latency, median (decode pace) | 15.9ms/tokenmedian over valid requests · lab_repeated · evidence | 18.7ms/tokenmedian over valid requests · lab_repeated · evidence |
| Strict-JSON task success | 100.0% of requestsshare of all issued requests · lab_repeated · evidence | 100.0% of requestsshare of all issued requests · lab_repeated · evidence |
How to read this
On the workloads these gates can see, the bigger artifact bought nothing: task success is identical, and Q8_0 decodes slower. If your workload resembles these — everyday chat and strict-JSON extraction — the memory is better spent on context.
Limits
Quality gates here are format/JSON/repetition classes — a quant difference could surface on harder reasoning or multilingual generation these workloads do not probe. Negative result, scoped to what was measured.
Answering this question somewhere? Paste the table
Markdown, renders on GitHub / Reddit / forums. Attribution line included — the data is CC BY 4.0, take it.