DeepSeek-V4.1-Flash-1017 on one 128 GB Mac: a ds4 Q2 GGUF is out, but vision waits on PR #1225

When DeepSeek put DeepSeek-V4.1-Flash-1017 on Hugging Face, the smallest community quants we listed in our 1017 open-weights guide started at 246 GiB and needed a llama.cpp branch. There is now a second local route: DwarfStar (ds4), antirez’s MIT-licensed DeepSeek inference engine for Metal, CUDA and ROCm. A ds4-format GGUF of 1017 went up on October 18, 2026, and it targets a single 128 GB Mac.

What was published

The repo is nazeshinjite/DeepSeek-V4.1-Flash-1017-ds4-GGUF (MIT). It’s a community build, not an official antirez or DeepSeek release. From the model card:

File	Size	Main weights
DeepSeek-V4.1-Flash-1017-Q2.gguf	340.60 GiB	151.77 GiB
DeepSeek-V4.1-Flash-1017-Vision.gguf	0.90 GiB	separate vision encoder

The other 188.83 GiB of the language file is DeepSeek’s native FP8 Engram tables. ds4 reads the rows it needs straight from disk and never loads the whole table, so the file belongs on a fast local SSD.

The card says it uses ds4’s stock --quant q2 recipe: IQ2_XXS routed gate/up experts, Q2_K down experts, Q8_0 attention, shared experts and output, and the Engram tables copied byte for byte. Tensor layout and file size (365,713,686,528 bytes) match antirez’s original DeepSeek-V4.1-Flash-Q2.gguf exactly. The one difference is the importance matrix, which was collected on the 1017 weights. The card notes that 32 of the 15,360 routed experts never got a calibration token and were quantized without importance weighting.

Running the text model today

The card says the language model loads in current ds4 with no changes:

hf download nazeshinjite/DeepSeek-V4.1-Flash-1017-ds4-GGUF DeepSeek-V4.1-Flash-1017-Q2.gguf --local-dir gguf
./ds4 -m gguf/DeepSeek-V4.1-Flash-1017-Q2.gguf --ssd-streaming --ctx 32768

Memory options are the same as for the first V4.1 checkpoint in the ds4 model guide. That means SSD streaming on one 128 GB Mac, resident tensor parallelism across two 128 GB Macs or DGX Sparks (each rank holds about 81 GiB of main weights), or full residency on a 256 GB or larger Mac. The guide also lists CUDA text inference with Q2 SSD streaming on one Spark, and says vision requires Metal. On speed, the quant’s author says decode “can reach up to 20tps” on an M5 Max with expert caching. That’s the author’s own figure; the PR itself states no speed benchmarks were run.

Vision: don’t reuse the old encoder

ds4 PR #1225, opened by the same author, was still open with no reviews when we checked. Current ds4 checks a V4.1 vision GGUF against one fixed revision of the first checkpoint (df42c109). The PR’s test table shows what that means on main:

• A correctly built 1017 vision file is rejected.
• The first checkpoint’s vision file loads next to a 1017 model without complaint, even though its vision tower was trained with different language weights.

The PR changes the check so the vision file must match its language model’s revision. It also adds a --source-revision flag to deepseek41_vision.py. Until it merges, using 1017 with images means building ds4 with the PR applied (for example git fetch origin pull/1225/head:pr-1225 in your clone), then passing --vision gguf/DeepSeek-V4.1-Flash-1017-Vision.gguf. The PR was tested on an M5 Max with 128 GB; two-machine TP was not tested.

Should you switch?

If you already run the first V4.1 checkpoint in ds4, the download is the same size and the flags don’t change. DeepSeek’s own 1017 update notes cover the benchmark gains. If you don’t have a 128 GB Mac or a Spark, the deepseek-flash API in our V4 Flash tutorial is still the cheaper way to try it.

Sources: nazeshinjite ds4 GGUF card, ds4 PR #1225, ds4 model guide, DeepSeek-V4.1-Flash-1017 model card, antirez V4.1 Flash GGUF.