What this is:        redux EDGE variant -- rides just under 4 GB
Trial:               Trial 24 of redux  (auto-picked Pareto)
Quant scheme:
  tier1 (8-bit affine, group 64) : mlp.down_proj on all 36 layers
  tier2 (6-bit affine, group 64) : self_attn.o_proj on layers 12-35
                                    (projection_pass-modified only)
  rest                            : 2-bit ternary (zero-preserving)
Caveat:              On greedy decoding (temp=0), produced degenerate output
                     ", , , , ," on several prompts. At higher temperatures
                     the model is more usable but FRAGILE on simple prompts.
                     Some Lucifer-style persona prompts produced more aggressive
                     in-character output than the safer mixed-d6 variant.
                     User wanted to test it themselves; here it is.