altic-dev/Fluid-1-Mini-2B-MLX-6bit

Fluid-1 Mini (v23) — FluidDecode Q6 + DFlash v3

2B dictation model (Qwen3.5 base, step 7314) quantized to affine Q6 group 64 for the FluidDecode runtime, with the v23 five-layer block-16 DFlash drafter attached, served at block 8 (taps [2, 7, 12, 17, 21], full 98k draft vocabulary, causal block attention). Output is byte-identical to plain greedy decoding. Requires FluidIntelligence with causal DFlash block attention (September 2026 or later).