DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF

Uncensored GGUF variant of AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF, converted to ISTA-DASLab GSQ-RCO IQ1_M and IQ2_XXS shards with a shared Engram IQ4_NL file. It is intended for one AMD Strix Halo machine (gfx1151, 128 GB unified memory) using the Gufo fork: https://github.com/franzmoca/gufo/tree/feat/ds41

Usage notes: https://github.com/franzmoca/gufo/blob/feat%2Fds41/README_DEEPSEEK.md

This is an uncensored model. Review the upstream model card and use it responsibly.

Files

Drafter credit: the DSpark drafter was contributed by Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF.

The 00006 shard of both IQ1_M and IQ2_XXS points to the same Engram IQ4_NL file.

Download

hf download dineshr93/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF \
  --repo-type model \
  --local-dir models/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF

Hardware

AMD Ryzen AI Max+ 395 (128 GB), AMD Radeon 8060S (gfx1151), Ubuntu 26.04 LTS.

Run with Gufo

export MODEL='models/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00006.gguf'

./result/bin/gufo serve llm \
  --model "$MODEL" \
  --context 162000 \
  --served-model-name=local-model \
  --host 0.0.0.0 \
  --port 8888 \
  --stream-experts

Memory note: this sharded model with streamed experts runs at extreme memory pressure on 128 GB. Inference works reliably only when no other GPU services are running (no other model servers, containers, or concurrent model loads).

Add vision support with the matching mmproj file:

--mmproj models/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF/mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf

The mmproj file is included in this repository.

For DSpark speculative decoding, append the repo's drafter (contributed by Lucebox/DeepSeek-V4.1-Flash-1017-ROCMFP-GGUF):

--speculative dspark \
--dspark-model models/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF/DeepSeek-V4.1-Flash-1017-DSpark-draft-Q2K-Q4K.gguf

Benchmark

Run llama-benchy against the served local-model alias:

uvx llama-benchy --base-url http://0.0.0.0:8888/v1 --model local-model

Measured on the AMD Strix Halo host above:

These results were measured without the DSpark drafter. With the DSpark drafter, generation is about 30 tok/s.

Real server timing from the already-running API (no drafter):

The llama-benchy pp2048 value is derived from the first response event; Gufo streams a usage event before content, so it is not a true prompt-processing rate. For real prefill rates, use Gufo's own gufo bench or server timing. The documented Gufo IQ1_M at 2048-token prefill is about 1,243.84 tok/s, and its single-user tg is about 24.71 tok/s at depth 0.