DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-GGUF (Zero-LoRA, llama.cpp b11870+ & Strata 6-Shard Layout)

This repository provides a standalone, weight-baked abliterated (uncensored) GGUF build of ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF (IQ2_XXS).

Update (October 2026): Upgraded to the official llama.cpp (b11870+ / v0.8.0+) & Strata 6-shard deepseek41 layout (00001-of-00006.gguf through 00006-of-00006.gguf) with embedded Jinja chat template, deepseek41.attention.kv_source_layers metadata, and mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf vision support.

Key Highlights

• Zero LoRA Required at Runtime: All 86 abliterated residual-writing tensors are baked directly into the quantized GGUF weights (blk.0..39.attn_output_b.weight upgraded to Q6_K, blk.0..39.ffn_down_shexp.weight, and 6 ffn_down_exps.weight layers).
• Native llama.cpp (b11870+ / b12017+) & Strata Support: Loads directly in mainline llama.cpp (llama-cli / llama-server) and strata-dsv41 without external adapters.
• Same Size as Original (138.58 GB vs. 138.24 GB): Adds only +0.34 GB (+0.25%) across the entire 5-shard backbone.
• 100% GSQ-RCO Intelligence Preserved: All 28 1-bit IQ1_M GSQ-calibrated routed-expert layers and the 110.6 GB IQ4_NL Engram lookup table remain 100% untouched byte-for-byte.

Empirical Verification & llama-perplexity KL-Divergence Audit (b12017)

Verified end-to-end against stock ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF (IQ2_XXS) across all 40 layers and all 129,280 vocabulary logits using llama-perplexity --kl-divergence (build b12017):

Raw llama-perplexity --kl-divergence Output (b12017)

chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
   1       2.4857 ±    0.5017       0.05455 ±    0.01477       0.00955 ±    0.00192     2.747 ±  0.500 %    98.425 ±  1.109 %

====== Perplexity statistics ======
Mean PPL(Q)                   :   2.485713 ±   0.501683
Mean PPL(base)                :   2.353761 ±   0.452226
Cor(ln(PPL(Q)), ln(PPL(base))):  99.84%
Mean ln(PPL(Q)/PPL(base))     :   0.054545 ±   0.014768
Mean PPL(Q)/PPL(base)         :   1.056060 ±   0.015596
Mean PPL(Q)-PPL(base)         :   0.131953 ±   0.056319

====== KL divergence statistics ======
Mean    KLD:   0.009546 ±   0.001921
90.0%   KLD:   0.029794
Median  KLD:   0.000421
10.0%   KLD:   0.000010
 5.0%   KLD:   0.000006
 1.0%   KLD:   0.000002
Minimum KLD:   0.000001

How to Download & Run (No LoRA Needed — llama.cpp & Strata)

Important: Both llama.cpp (b11870+) and Strata open the shared DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf (110.6 GB) lookup file as Shard 6 (...-00006-of-00006.gguf). After downloading, create a symlink (Linux/macOS) or hardlink (Windows) so Shard 6 is linked automatically:

# 1. Download all 5 backbone shards + the Engram lookup file + mmproj
hf download Abiray/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-GGUF \
  --local-dir ./DeepSeek-V4.1-Uncensored

cd ./DeepSeek-V4.1-Uncensored

# 2A. Link the Engram file as Shard 6 (Linux / macOS):
ln -s DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00006-of-00006.gguf

# 2B. Or on Windows (Command Prompt / PowerShell):
cmd /c mklink /H DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00006-of-00006.gguf DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf

Option 1: Run with llama.cpp (b11870 / v0.8.0 or newer) — Recommended

# CLI Inference
llama-cli -m DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf \
  -ngl 99 --reasoning on --reasoning-effort 50 \
  -p "Explain mixed-precision quantization."

# Server Inference (with Vision Projector)
llama-server -m DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf \
  --mmproj mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf \
  --jinja -ngl 99 --reasoning on --reasoning-effort 50 --port 8080

Option 2: Run with Strata (1x 12 GB+ NVIDIA GPU)

Note for Strata 0.2.8 (build 0.1.54): Always pass --family deepseek41 --model ist-das-2.00bpw and the directory path (--model-path ./DeepSeek-V4.1-Uncensored). Because this uncensored build upgrades blk.0..39.attn_output_b.weight to Q6_K (138.58 GB vs. stock 138.24 GB), if Strata 0.1.54's hardcoded stock file-size check in setup.py reports "this is not a model size Strata knows", either connect Strata to llama-server above (http://127.0.0.1:8080) as a Custom Model API, or build the pack directly via Strata's bundled tools/dsv41/pack_gguf.py.

curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh | sh
strata launch --family deepseek41 --model ist-das-2.00bpw --model-path ./DeepSeek-V4.1-Uncensored --full-quality

Upgrading Existing Pre-Oct 27 4-Shard Downloads Locally (No Re-Download Needed)

If you previously downloaded the 00001-of-00004.gguf layout, you can convert your local folder into the 6-shard 00001-of-00006.gguf layout in place using the included patched convert_ds41_upstream.py script:

python convert_ds41_upstream.py /path/to/your/model/folder --delete-old

Shard Integrity & SHA256 Checksums

Acknowledgements & Credits

This release builds directly upon research, quantization breakthroughs, and tooling from across the open-source AI ecosystem:

• DeepSeek AI — For open-sourcing the base DeepSeek-V4.1 architecture and foundation weights.
• ISTA-DASLab — For developing the GSQ-RCO quantization pipeline, preserving the 1-bit IQ1_M routed experts with zero degradation on AIME25/GPQA-D, and providing the shared DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf lookup backbone.
• AtomicChat — For identifying and isolating the orthogonal refusal directions (EN+RU) that enabled weight-baked abliteration without requiring runtime LoRA adapters.