DeepSeek-V4-Flash · MLX - a pipenetwork Collection

pipenetwork 's Collections

Inkling-Small · MLX

DeepSeek-V4-Flash · MLX

Nemotron TwoTower · MLX

Qwen3.6-35B-A3B — MLX

Ornith-1.0-397B — MLX

GLM-5.2 REAP (expert-pruned)

Rio-3.1-Open-30B MLX

Gemma-4-26B-A4B-it MLX

Gemma-4-31B-it MLX

Kimi-K2.7-Code MLX

Holo-3.1 MLX (computer-use)

Frog (SWE/debugging) MLX

Nemotron-3 MLX (Apple Silicon)

DeepSeek-V4-Flash · MLX

updated about 9 hours ago

304B MoE on Apple Silicon. deepseek_v4 is in no released runtime — ported from scratch. Take mixed-4_8bit. github.com/PipeNetwork/deepseek-v4-mlx

pipenetwork/DeepSeek-V4-Flash-MLX-8bit

Text Generation • 81B • Updated about 12 hours ago • 27

Note 304 GB · ppl 5.9878 · highest fidelity
pipenetwork/DeepSeek-V4-Flash-MLX-6bit

Text Generation • 64B • Updated about 12 hours ago • 23

Note 233 GB · ppl 6.0260 · +0.6% vs 8-bit for 70 GB less
pipenetwork/DeepSeek-V4-Flash-MLX-mixed-4_8bit

Text Generation • 47B • Updated about 12 hours ago • 32

Note RECOMMENDED · 165 GB · ppl 6.1262 · 4-bit routed experts, 8-bit everything else. Experts are 98.1% of params, so raising the other 2% costs 3 GB and recovers 56% of the gap to 8-bit.
pipenetwork/DeepSeek-V4-Flash-MLX-4bit

Text Generation • 46B • Updated about 12 hours ago • 30

Note 162 GB · ppl 6.3005 · superseded by mixed-4_8bit — 3 GB more, 2.8% better
pipenetwork/DeepSeek-V4-Flash-MLX-REAP25

Text Generation • 36B • Updated about 12 hours ago • 31

Note 129 GB · ppl 6.4025 · 25% of experts pruned; statistically indistinguishable from 4-bit at 33 GB less
pipenetwork/DeepSeek-V4-Flash-MLX-REAP37

Text Generation • 32B • Updated about 12 hours ago • 33

Note 111 GB · ppl 6.9634 · 37% pruned; fits a 128 GB Mac
pipenetwork/DeepSeek-V4-Flash-MLX-REAP50

Text Generation • 26B • Updated about 12 hours ago • 32

Note 93 GB · ppl 7.7520 · 50% pruned; smallest build