The main readout is now on the repo front page. This page keeps the caveats in one place.

What changed the outcome most

The biggest divider was context window, not f16 vs q8 KV.

  • The 65k profiles worked inside the 65k envelope.
  • The 65k profiles could not handle the >65k needle prompt.
  • The successful 128k profiles passed both needle tests.

q8 KV finding

This exact suite did not show an accuracy loss from q8 KV among the successful 128k profiles.

These tied:

  • bartowski-128k-f16
  • bartowski-128k-q8
  • unsloth-128k-q8

That does not prove q8 is always equal to f16. It only says this task-specific suite did not expose a deficit.

Unsloth 128k f16 finding

unsloth-128k-f16 loaded locally, but the two long-context cases timed out. The useful wording is “not viable on this machine as configured,” not “answered incorrectly.”

Practical recommendation

For Robert’s local Codex + Qwen sidecar workflow, the best tested operating tier is:

  1. unsloth-128k-q8
  2. bartowski-128k-q8
  3. bartowski-128k-f16

This is a workflow-specific result, not a universal model ranking.