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

MLX (Apple Silicon) build of DeepSeek-V4.1-Flash — 754.6B parameters: 40 layers x 384 routed experts, MLA with cross-layer KV-cache sharing (4 compressor-owning layers serve all 40), a two-layer engram hashed n-gram embedding whose tables alone are 196.6B parameters (~40% of the checkpoint), staggered Sinkhorn hyper-connections, and per-layer attention sinks — quantized to 4-bit experts / 8-bit attention&shared / 8-bit engram.

These files are modified: dequantized from the FP8/FP4 release (bit-exact decode of the 32x32-block ue8m0 fp8 and per-32 fp4 packing) and re-quantized; the architecture is unchanged. The 3 multi-token-prediction layers (DSpark markov/confidence heads) are not included; the vision tower and aligner are carried unmodified but the runtime is text-only.

Runtime

deepseek_v41 exists in no runtime — not transformers, not mlx-lm, not mlx-vlm. This checkpoint loads only through the port:

git clone https://github.com/PipeNetwork/deepseek-v41-mlx && cd deepseek-v41-mlx && pip install -r requirements.txt
python scripts/smoke_generate.py /path/to/DeepSeek-V4.1-Flash-MLX-mixed-4_8bit

from deepseek_v41_mlx.load import load
model, tokenizer = load("/path/to/DeepSeek-V4.1-Flash-MLX-mixed-4_8bit")

The port was validated against DeepSeek's own inference/model.py (the only reference): fp32 tiny-config parity 1e-6 across prefill / cached decode / chunked prefill, the three QAT fake-quant ops bit-exact, and negative controls proving the fragile paths are load-bearing (rope-inverse 0.84, attention sinks 0.65, cross-layer sharing 0.56 logit shift when broken). Strict loading reports zero missing / zero unexpected tensors. One reference decode bug was found and documented (odd-step indexer reads the wrong layer's keys; 0.67 logit shift — the port uses the owner's cache): see docs/upstream-notes.md in the repo.

Size and what is quantized

427.6 GB on disk. RAM: 512 GB Mac (tight: 427 GB resident).

group	share of parameters	this build
routed experts (40 x 384, `w1/w2/w3`)	543.6B (72%)	4-bit, group 64
engram tables (2 x [384,006,168 x 256])	196.6B (26%)	8-bit, group 64
attention (MLA), shared experts, embeddings, `head`	~14B	8-bit, group 64
`wo_a` (block-diagonal output LoRA), hyper-connections, sinks, router biases, compressor, indexer keys, norms	—	unquantized (bf16/fp32)

Quality

Per-layer divergence ladder vs the bf16-dequantized reference — every one of the 40 decoder layers run on identical inputs (16,384 tokens of wikitext-2), teacher-forced and free-running, with the ladder's arithmetic asserted bit-identical to this converter's:

recipe	teacher-forced (mean)	free-running (final layer)	cosine (final)
8bit	0.0084	0.1243	0.9910
6bit	0.0177	0.1393	0.9886
4bit-engram4	0.0579	0.2714	0.9634
mixed 4/8, engram as shipped (fp8/ue8m0)	0.0335	0.1948	0.9800
mixed 4/8, engram 6-bit (this set's 1 TB build)	0.0335	0.1945	0.9801
mixed 4/8, engram 4-bit (this set's 512 GB build)	0.0342	0.2090	0.9775

The engram finding that shapes this set: 6-bit engram is indistinguishable from the shipped fp8 tables (free-running 0.1945 vs 0.1948) while 4-bit engram costs +7.3% free-running — but the engram-6 build is 477 GB and a 512 GiB machine cannot run it, so the engram-4 build is the one that fits and the engram-6 build serves 1 TB machines.

Perplexity (wikitext-2 test, 286,580 tokens in 140 windows of 2048, through this runtime): 2.8963 [2.7103, 3.0933]. Greedy generation is coherent (collapse check).