DeepSeek-V4.1 Flash 1017 MXFP4 GGUF for DwarfStar
GGUF files for DeepSeek V4.1 Flash 1017 in DwarfStar (DS4). They work with
upstream main and with the kernelpool/ds4:ds41-perf
branch of antirez/ds4#1073, which speeds up V4.1 decoding and
prefill on Metal. These are DS4 files, not llama.cpp GGUFs: llama.cpp does not load DS4's deepseek41 layout.
The files come from deepseek-ai/DeepSeek-V4.1-Flash-1017;
files for the original September release are in
kernelpool/DeepSeek-V4.1-Flash-MXFP4-GGUF.
Keep the main model, the drafter and the vision encoder from the same release: the two checkpoints have
different weights, including the DSpark stages and the vision tower.
| File | Size | Content |
|---|---|---|
DeepSeek-V4.1-Flash-1017-MXFP4.gguf.part1, .part2 |
467.2 GiB | Main model, deepseek41 layout, in two parts: routed experts are the checkpoint's own MXFP4 blocks repacked without changing their values; Q8_0 attention, shared experts and output head; F16 embeddings; the two Engram tables in the checkpoint's FP8 encoding (188.8 GiB, read from disk) |
DeepSeek-V4.1-Flash-1017-DSpark.gguf |
8.0 GiB | DSpark drafter for speculative decoding, deepseek41-dspark layout: native MXFP4 experts, dense weights expanded losslessly to F16 or F32 |
DeepSeek-V4.1-Flash-1017-Vision.gguf |
0.9 GiB | Vision encoder for image input, deepseek4-vision layout: the checkpoint's vision tower and aligner in BF16 |
Run
Join the two parts first (the parts can be deleted afterwards):
cat DeepSeek-V4.1-Flash-1017-MXFP4.gguf.part1 DeepSeek-V4.1-Flash-1017-MXFP4.gguf.part2 > DeepSeek-V4.1-Flash-1017-MXFP4.gguf
Build either tree, then run:
git clone https://github.com/antirez/ds4 && cd ds4 && make # upstream
git clone -b ds41-perf https://github.com/kernelpool/ds4 && cd ds4 && make # antirez/ds4#1073
./ds4 -m DeepSeek-V4.1-Flash-1017-MXFP4.gguf
./ds4 -m DeepSeek-V4.1-Flash-1017-MXFP4.gguf --dspark --mtp-model DeepSeek-V4.1-Flash-1017-DSpark.gguf
./ds4 -m DeepSeek-V4.1-Flash-1017-MXFP4.gguf --vision DeepSeek-V4.1-Flash-1017-Vision.gguf
./ds4-server -m DeepSeek-V4.1-Flash-1017-MXFP4.gguf --dspark --mtp-model DeepSeek-V4.1-Flash-1017-DSpark.gguf
The main model keeps 278 GiB of weights resident and reads the Engram tables from disk; the drafter adds
8 GiB. Tested on a 512 GB Mac Studio M3 Ultra. DSpark is optional and does not help every prompt; see
speculative decoding. With
--vision, /read FILE in ./ds4 submits a PNG or JPEG.
Conversion
Written from deepseek-ai/DeepSeek-V4.1-Flash-1017
at revision 1bae0e1088993b9b82de08f59de5be5a34e42170, which the main model and drafter record in
general.source.revision. gguf-tools/deepseek41_vision.py always records the September revision
df42c109, so the vision encoder's metadata names that revision although it holds the 1017 weights. All
three converters write the original repo's URL into general.source.url. The MXFP4 mode of
gguf-tools/deepseek41_quantize.py is in the ds41-perf branch (upstream's converter writes Q2 and Q4);
gguf-tools/deepseek41_dspark.py and gguf-tools/deepseek41_vision.py are in both trees.
make -C gguf-tools libds4quants.dylib
python3 gguf-tools/deepseek41_quantize.py --hf models/DeepSeek-V4.1-Flash-1017 \
--source-revision 1bae0e1088993b9b82de08f59de5be5a34e42170 \
--quant mxfp4 --out DeepSeek-V4.1-Flash-1017-MXFP4.gguf
python3 gguf-tools/deepseek41_dspark.py --hf models/DeepSeek-V4.1-Flash-1017 \
--source-revision 1bae0e1088993b9b82de08f59de5be5a34e42170 \
--out DeepSeek-V4.1-Flash-1017-DSpark.gguf
python3 gguf-tools/deepseek41_vision.py --hf models/DeepSeek-V4.1-Flash-1017 \
--out DeepSeek-V4.1-Flash-1017-Vision.gguf
Quality
Scored with gguf-tools/quality-testing/score_official on the six tracked V4.1 Flash fixture sets
(gguf-tools/quality-testing/deepseek-v4.1-flash-*, 256 cases): upstream main and the ds41-perf
branch give identical results with this file. With the vision encoder, both trees also give identical
output on the image fixtures in tests/vision-fixtures/glm53 (text.png, earth.jpg).