DeepSeek-V4.1-Flash-1017 — sidecars (vision tower, DSpark drafter, llama-layout skeleton)
The three small companion files the ds4 / ds4-server runtime loads next to a lecore-quantized DeepSeek-V4.1-Flash-1017 build. None of these is a standalone model — the weights themselves live in the 99–107 GB VQ GGUFs listed at the bottom, and only the ds4 runtime can load those.
| file | size | what it is | how it is loaded |
|---|---|---|---|
DeepSeek-V4.1-Flash.vision.gguf |
0.90 GiB | Vision tower sidecar: 32 ViT blocks, width 1024, patch 14, downsample 3, ≤1024 image tokens per image. Architecture deepseek4, sidecar=vision, 306 tensors. Emitted from the unquantized model; shared by every V4.1-Flash build, abliterated or not. |
ds4-server --vision-gguf <file> (already the default in run_ds41.ps1) |
DeepSeek-V4.1-Flash-1017-abliterated-pqhaz-lecore-vq11cal.dspark.llama.gguf |
5.05 MiB | DSpark speculative drafter: 3 draft towers × 128 experts, top-3, block 5, target layers 37/38/39, confidence head. Weights are the pqhaz-abliterated draft towers, VQ 8×1024 codewords. | picked up from the model directory next to the main GGUF (drafter armed automatically) |
DeepSeek-V4.1-Flash-1017-abliterated-pqhaz-lecore-vq11cal.llama.gguf |
605.8 MiB | Dense-skeleton emit in llama.cpp GGUF layout: all non-expert tensors (attention q_a/q_b/kv/output_a/output_b, shared-expert FFN, indexer, MTP main_proj, token embedding, output head) with the full deepseek41.* metadata, 922 tensors. The routed experts are not in it — they stay in the 103 GB VQ file. |
metadata/type-pass sidecar; not a loadable model on its own |
Use
git clone https://huggingface.co/staccs/ds4-winhip # patched ds4 runtime + build scripts
cd ds4-winhip
.\hip\win\run_ds41.ps1 -Gguf D:\models\ds41\DeepSeek-V4.1-Flash-1017-abliterated-pqhaz-lecore-vq11cal.gguf `
-Engram D:\models\ds41\deepseek-engram
run_ds41.ps1 points --vision-gguf at DeepSeek-V4.1-Flash.vision.gguf inside the same model
directory, so drop this file next to the main GGUF (or pass -Extra "--vision-gguf,<path>").
Measured DSpark acceptance on the pqhaz build (Radeon 8060S, 128 GB unified): 1.0–3.3 accepted tokens per 5-token verification block depending on prompt mix; with any repetition penalty set the drafter is disabled by the runtime and you get plain decode. Vision encoder loads additively — the image tower is ~1 GiB of weights and does not touch the text path.
Engram n-gram tables
The main GGUFs do not contain DeepSeek's Engram tables. They are the last two shards of the upstream repo, used unchanged (~203 GB, read from disk per token):
- model-00047-of-00048.safetensors — 101.5 GB
- model-00048-of-00048.safetensors — 101.5 GB
Drop both in one directory and pass --engram-dir <dir> (or -Engram <dir>); the server maps
them, it does not copy them into GPU memory.
Main weights
| build | size | PPL (WikiText-2, 2k) | top-1 vs source |
|---|---|---|---|
| pqhaz abliterated, vq11cal | 106.8 GB | 2.78 | 80.4% |
| dealignai abliterated, vq11cal | 106.8 GB | 2.82 | 80.2% |
| mrfakename abliterated, vq11cal | 106.8 GB | 2.80 | 80.3% |
| rumplestilzken abliterated, vq11cal | 106.8 GB | 2.82 | 80.5% |
| base model, vq10sh12 | 99.1 GB | 3.615 | 74.5% |
License
MIT, following the base model. The vision tower and Engram tables are DeepSeek's own weights, redistributed unchanged.