DeepSeek-V4-Flash-MLX-REAP25

DeepSeek-V4-Flash-MLX-REAP25

MLX (Apple Silicon) conversion of deepseek-ai/DeepSeek-V4-Flash-0731, built from mixed-4_8bit (4-bit routed experts, 8-bit elsewhere) with 25% of routed experts pruned by REAP saliency.

Code / loader: github.com/PipeNetwork/deepseek-v4-mlx

DeepSeek-V4-Flash is a 304B-parameter sparse-MoE model. Its deepseek_v4 architecture is not implemented in any released runtime — not transformers, and not mlx-lm (which carries deepseek, _v2, _v3, _v32). This repo bundles a from-scratch MLX port under deepseek_v4_mlx/, validated against DeepSeek's own reference implementation.

Builds

Perplexity is teacher-forced over wikitext-2-raw-v1 test, 199,485 tokens in 195 windows of 1024 — identical windows for every build. Differences are established by a paired bootstrap over shared windows, which removes window-difficulty variance: independent intervals overlap almost entirely and settle nothing, while the paired interval is ~10x tighter. Every gap quoted here is significant at 95%.

Why mixed-4_8bit is the recommended build

The routed experts are 98.1% of all quantized parameters, and they ship natively as FP4. Keeping them at 4 bits while raising everything else — attention, embeddings, shared experts, about 5B parameters — to 8 bits costs 3 GB and recovers 56% of the gap between 4-bit and 8-bit. Per parameter, the non-expert weights are roughly 50x more quantization-sensitive.

Two variants were built, measured, and not published:

• group-size 32 (matching FP4's native 32-value scale blocks) is 0.5% worse than group 64 and 18 GB larger. Finer scaling improves raw tensor reconstruction but not prediction — the mismatch is that FP4's levels are non-uniform while affine's are evenly spaced, which no group size fixes.
• 3-bit experts collapse the model entirely (perplexity 1.4e7, degenerate token loops). 4-bit sits just above a cliff for this checkpoint, not in the middle of a gentle curve.

REAP-pruned builds

REAP (Cerebras, arXiv:2510.13999) drops the lowest-saliency routed experts, where saliency is the mean of router_weight x ||expert_output|| over the tokens routed to that expert — its actual contribution, not how often it fires. The router renormalizes over survivors.

Only 40 of the 43 layers are pruned. The first three route through a fixed tid2eid token-id table with no router to renormalize, so dropping an expert there would leave table entries pointing at nothing. They keep all 256, and the config records two expert counts which the loader resolves per layer.

Saliency was profiled over 150k tokens of English prose, ten other languages and real code — deliberately disjoint from the evaluation set, verified by 32-gram overlap (0 of 4687), so pruning decisions are not fit to the text they are judged on.

REAP25 is the one to take — its perplexity is statistically indistinguishable from plain 4-bit (paired ratio 1.016, CI [0.998, 1.038], crossing 1.0) at 33 GB less.

Note how poorly saliency retention predicts damage: 96.2% retained still costs 4.5%, and the last 12 points of retention cost 22 points of perplexity. Treat it as a screen for whether pruning is viable, not as an estimate of what it will cost.

⚠️ Loading requires the bundled loader

pip install mlx mlx-lm transformers

from deepseek_v4_mlx.load import load
from deepseek_v4_mlx.generate import greedy_generate, load_tokenizer

model, args = load("/path/to/this/repo")
tok = load_tokenizer("/path/to/this/repo")   # bundled — no download
ids = tok("The capital of France is")["input_ids"]
print(tok.decode(greedy_generate(model, args, ids, max_new_tokens=32)))

The tokenizer is bundled in this repo, so nothing is fetched at runtime. (The base repo shipped no tokenizer at first — only encoding/encoding_dsv4.py, a prompt-string renderer; DeepSeek has since published one. DeepSeek-V3's tokenizer, used here in the interim, produces identical ids for text and identical special-token ids, so builds made either way agree.)

What this port implements

The source checkpoint is doubly pre-quantized — FP8 e4m3 with e8m0 128x128 block scales for most matmuls, FP4 e2m1 packed two-per-byte for routed experts — so conversion dequantizes both formats before requantizing. Both decode paths are bit-exact against an independent implementation.

Dropped: the multi-token-prediction head and DSpark (speculative decode over 5-token blocks in the last three layers). Both hang off the reference's forward_spec path and are unused for ordinary generation.

wo_a is kept unquantized: it is a block-diagonal projection the model reshapes by group, and a packed quantized weight cannot be reshaped that way. Costs ~2% of build size.

License: MIT (inherits the base model).