spoomplesmaxx-mockingbird-36B-i1-GGUF

spoomplesmaxx-mockingbird-36B — i1-GGUF (weighted/imatrix)

Weighted/imatrix GGUF quants of spoomplesmaxx-mockingbird-36B. The importance matrix was computed on a stratified sample of the model's own training corpus — all ten lanes, rendered in the exact chat template the model serves with — not a generic calibration set. At 3–4 bit these should beat the static quants noticeably; at Q5 the difference fades.

The seed-native chat template is embedded in the GGUF metadata.

Sampling — read this part

temperature 1.0 · top_p 0.9 · repeat_penalty 1.0 (OFF)

⚠ Never use repetition, presence, or frequency penalties. The template ends every message with <seed:eos>; context-wide penalties suppress that token, the model stops ending its turns, and generation degenerates into the base model's untrained Chinese vocabulary. Many frontend presets default repeat_penalty to 1.05–1.1 — set it back to 1.0. Use DRY or XTC if you want extra anti-repetition; both leave special tokens alone.

Usable temperature window is ~0.95–1.05: lower loops verbatim, higher frays. Full details on the main model card.

mimids 01 · Apache 2.0