fly88oj/decidex-draft-0.6b

Decidex draft 0.6B — the MTP/speculative-decoding companion

Qwen3-0.6B + LoRA (r=16) distilled on the same decision corpus, merged and exported as GGUF Q8_0. Its letter distribution matches the 8B targets, so it makes a high-acceptance draft for llama.cpp speculative decoding:

llama-server -m decidex-core-8b-Q4_K_M.gguf \
  --model-draft decidex-draft-0.6b-Q8_0.gguf \
  --spec-draft-n-max 4 -ngl 99

Note: speculative decoding accelerates generation; the decision readout itself generates zero tokens. See GGUF-USAGE.md in the Decidex repo. Not affiliated with TypeSafe AI.