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.
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-1017-MLX-mixed-4_8bit
from deepseek_v41_mlx.load import load
model, tokenizer = load("/path/to/DeepSeek-V4.1-Flash-1017-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.
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) |
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) |
|---|---|---|---|
| mixed 4/8, engram as shipped (fp8/ue8m0) | 0.0337 | 0.1947 | 0.9803 |
| mixed 4/8, engram 6-bit (this set's 1 TB build) | 0.0338 | 0.1928 | 0.9804 |
| mixed 4/8, engram 4-bit (this set's 512 GB build) | 0.0344 | 0.2126 | 0.9769 |
The engram finding that shapes this set: 6-bit engram is indistinguishable from the shipped fp8 tables (free-running 0.1928 vs 0.1947) while 4-bit engram costs +9.2% 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): 2.9238 [2.7364, 3.1264]. Greedy generation is coherent (collapse check). Paired against the previous checkpoint's build on identical windows, perplexity is higher (worse): ratio x1.0095 [1.0050, 1.0139], worse on 93/140 windows. Held-out prose perplexity is not the axis this checkpoint was improved on — upstream reports large agentic, reasoning and vision gains — so read it as a conversion-fidelity check, not a verdict on the model.
MIT, as the upstream model. Port code: https://github.com/PipeNetwork/deepseek-v41-mlx.
model-layers-14-engram.safetensors 55 GBmodel-layers-01-engram.safetensors 55 GBmodel-layers-14.safetensors 8.0 GBmodel-layers-01.safetensors 8.0 GBmodel-layers-02.safetensors 7.9 GBmodel-layers-08.safetensors 7.9 GBmodel-layers-20.safetensors 7.9 GBmodel-layers-24.safetensors 7.9 GBmodel-layers-28.safetensors 7.9 GBmodel-layers-36.safetensors 7.9 GBmodel-layers-32.safetensors 7.9 GBmodel-layers-12.safetensors 7.9 GBmodel-layers-19.safetensors 7.9 GBmodel-layers-21.safetensors 7.9 GBmodel-layers-29.safetensors 7.9 GBmodel-layers-39.safetensors 7.9 GBmodel-layers-11.safetensors 7.9 GBmodel-layers-13.safetensors 7.9 GBmodel-layers-16.safetensors 7.9 GBmodel-layers-22.safetensors 7.9 GBmodel-layers-25.safetensors 7.9 GBmodel-layers-26.safetensors 7.9 GBmodel-layers-27.safetensors 7.9 GBmodel-layers-34.safetensors 7.9 GBmodel-layers-35.safetensors 7.9 GBmodel-layers-10.safetensors 7.9 GBmodel-layers-15.safetensors 7.9 GBmodel-layers-23.safetensors 7.9 GBmodel-layers-33.safetensors 7.9 GBmodel-layers-31.safetensors 7.9 GBmodel-layers-18.safetensors 7.9 GBmodel-layers-30.safetensors 7.9 GBmodel-layers-38.safetensors 7.9 GBmodel-layers-37.safetensors 7.9 GBmodel-layers-17.safetensors 7.9 GBmodel-layers-03.safetensors 7.9 GBmodel-layers-04.safetensors 7.9 GBmodel-layers-07.safetensors 7.9 GBmodel-layers-00.safetensors 7.9 GBmodel-layers-05.safetensors 7.9 GBmodel-layers-09.safetensors 7.9 GBmodel-layers-06.safetensors 7.9 GBmodel-top.safetensors 1.4 GBmodel-vision.safetensors 971 MB有文件？提交种子。
提交种子