DeepSeek-V4.1-Flash MLX - a pipenetwork Collection

pipenetwork 's Collections

Ternary-Bonsai-2 MLX

DeepSeek-V4.1-Flash MLX

GLM-5.3-Flash MLX

Qwen3.8-Flash-Next MLX

Qwen3.8-2.4T-A95B MLX (REAP)

Muse-Glimmer 30B MLX

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-V4.1-Flash MLX

updated Sep 10

Apple Silicon (MLX) builds of DeepSeek-V4.1-Flash (754B) with the only runtime that exists for it: github.com/PipeNetwork/deepseek-v41-mlx

pipenetwork/DeepSeek-V4.1-Flash-MLX-mixed-4_8bit

Text Generation • 379B • Updated Sep 10 • 2.61k • 4

Note 427 GB, 512 GB Mac. ppl 2.8963 on 287k tokens of wikitext-2. 4-bit experts, 8-bit attention, 4-bit engram tables (the 196B engram at 4-bit costs 7.3% free-running vs shipped fp8 - the price of fitting 512 GB).
pipenetwork/DeepSeek-V4.1-Flash-MLX-mixed-4_8bit-engram6

Text Generation • 428B • Updated Sep 10 • 1.6k • 2

Note 477 GB, 1 TB machines. 6-bit engram is indistinguishable from the shipped fp8 tables on the per-layer ladder (0.1945 vs 0.1948) - the quality pick where it fits. Cannot run on 512 GiB (measured, four load modes).
pipenetwork/DeepSeek-V4.1-Flash-REAP25-MLX-mixed-4_8bit

Text Generation • 311B • Updated Sep 10 • 888

Note 351 GB, 384 GB-class. 288 of 384 experts per layer. ppl 2.9777, paired x1.0281 [1.018, 1.039] vs unpruned - nearly free, the V4 family prunes well.
pipenetwork/DeepSeek-V4.1-Flash-REAP50-MLX-mixed-4_8bit

Text Generation • 243B • Updated Sep 10 • 846

Note 275 GB. 192 of 384 experts. ppl 3.3833, paired x1.1681 vs unpruned - visible degradation in generation style; the smallest tier, published with the numbers.