DeepSeek-V4-Flash — Vector Quantized (MLX) - a TheDrainFlorist Collection

TheDrainFlorist 's Collections

DeepSeek-V4-Flash — Vector Quantized (MLX)

GLM-5.3-Flash-VQ — Vector Quantized (MLX)

Qwen3.8-Flash-Next — Vector Quantized (MLX)

Qwen3.8-27B — Vector Quantized (MLX)

Gemma-4 — Vector Quantized (MLX)

Qwen3.6-35B-A3B — Vector Quantized (MLX)

Qwen3.5-397B-A17B — Vector Quantized (MLX)

DeepSeek-V4-Flash — Vector Quantized (MLX)

updated 30 days ago

Data-free VQ of DeepSeek-V4-Flash for Apple Silicon. One rung, 3.2 bpw, 6-bit MTP head. Runs on Knurlogic (macOS). Built with VQLab.

TheDrainFlorist/DeepSeek-V4-Flash-VQ-3.2bpw

Text Generation • 36B • Updated Oct 11 • 151

Note 106.4 GiB, d4/K2048 with K4096 on layers 24-35. KL 80.0 to DeepSeek's reference numerics — 21% less than uniform K2048 for +2.3 GiB. Two-Mac tensor split; 6-bit MTP head.
TheDrainFlorist/DeepSeek-V4-Flash-Vision-Exp-VQ-3.2bpw

Image-Text-to-Text • 37B • Updated Oct 11 • 256

Note 106.4 GiB text, d4/K2048 with K4096 on the middle layers. KL 93.2 to the exact teacher — 22% less than uniform K2048 for +2.3 GiB. Exact vision tower (+0.87 GiB); DSpark MTP head.