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

This repository provides a zero-loss uncensored (abliterated) build of ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF, created via surgical tensor transplantation from AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF.

Overview & How It Works

Standard abliteration methods often require dequantizing and re-quantizing model weights, which degrades the specialized Gumbel-Softmax Quantization (GSQ-RCO) calibration used by ISTA-DASLab.

To prevent any loss in coding intelligence:

Surgical Residual Transplant: Only the 82 dense residual-stream writers (attn_output_b, ffn_down_shexp, and engram_kv) were transplanted byte-for-byte from the uncensored donor.
100% Untouched Coding Experts: All routed coding expert layers (192 capability-pruned experts per layer) and router gates remain 100% byte-for-byte identical to the original ISTA-DASLab release.
Complete Standalone Repo: Includes the 72.1 GB patched IQ2_XXS backbone shards (00001 to 00005), the 110.6 GB shared Engram lookup table, vision projector (mmproj) in the root folder.

Quality & Benchmark Comparison (Original vs. Uncensored)

Because no routed coding experts were re-quantized, this model retains 100% of the coding and reasoning performance of the original model while completely eliminating safety refusals:

Benchmark / Metric	Original (`ISTA-DASLab`)	This Model (`Uncensored`)	Notes
Top-1 Token Output Agreement	`100.0%`	`99.80%`	Near-identical token generation on coding & logic prompts
40-Layer Residual Cosine Similarity	`1.0000`	`0.99999993`	Zero cumulative drift across all 40 transformer layers
Output Distribution Shift (Median KLD)	`0.0`	`6.94e-08 nats`	Statistically indistinguishable probability distribution
Coding & Reasoning Retention	`98.7%` LCB / `94.8%` SWE	`100% Parity`	All GSQ-RCO routed coding experts untouched
Safety Refusal Rate	`86.4%`	`0.0%`	Responds directly to security, raw-socket & restricted prompts

Quickstart & Running Instructions

Thanks to memory-mapped streaming (mmap), the 110.6 GB Engram lookup table streams directly from disk (1 row per token), allowing the full model to run comfortably inside 96 GB – 128 GB of system RAM (CPU) or GPU + RAM offload.

Option 1: `llama.cpp` (`b11870+`)

First, create a one-time symlink so llama.cpp recognizes the Engram file as Shard 6:

ln -s DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf \
  DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-Coder-IQ2_XXS-00006-of-00006.gguf

Then launch inference pointing to Shard 1:

./llama-cli \
  -m DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-Coder-IQ2_XXS-00001-of-00006.gguf \
  --jinja -c 16384 -n 2048 -t 16 --temp 0.2 \
  -p "Write a high-performance raw-socket TCP SYN scanner in Python using struct.pack."

Tip for Best Coding Accuracy: Keep <think> reasoning enabled (default) and use --temp 0.2 for complex systems or security code. If you want instant code output without the thinking monologue, add --reasoning-budget 0.

Option 2: `Strata` (`v0.2.5+`)

Strata automatically reads the Engram file from the root directory:

strata run --family deepseek41 \
  --model-dir . \
  --engram ./DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf

Credits

Base Coder Model: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF
DeepSeek AI — For open-sourcing the base DeepSeek-V4.1 architecture and foundation weights.
Uncensored Donor: AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF
Surgical Build & Verification: Abiray

Abiray/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-Coder-GGUF