license: mit
base_model:

• AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF
• ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
• deepseek-ai/DeepSeek-V4.1-Flash-1017
  base_model_relation: quantized
  pipeline_tag: text-generation
  library_name: gguf
  tags:
• deepseek
• deepseek41
• gguf
• gsq
• rco
• llama.cpp
• uncensored

DeepSeek-V4.1-Flash-1017 Uncensored GSQ-RCO: llama.cpp layout

This repo holds AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF
IQ2_XXS, rewritten into the layout that upstream llama.cpp loads (deepseek41 architecture, release
b11870 or newer).

AtomicChat publishes the 4-shard layout that Strata reads. Upstream llama.cpp
refuses that layout with array key not found: deepseek41.attention.kv_source_layers.

How these files were made

• Converter: ISTA-DASLab's convert_ds41_upstream.py at revision 65d53bc.
• Input: AtomicChat's shards at revision 4f1d2f7.
• Weights are unchanged. Quantized tensors are copied byte for byte, with no requantization. AtomicChat's
  uncensored edit (attention output, shared-expert down projection and Engram value) is preserved exactly. Every
  written tensor was read back and its SHA-256 compared with the source tensor.
• What changed is the same as for ISTA's own October 27 files: upstream metadata keys, engram_gate = q × k in
  F32, the vision router bias, the 2-D attn_output_a, and a metadata-only first shard. The chat template is ISTA's
  current chat_template.jinja.

Files

Running with llama.cpp

hf download nguyenthdat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF --local-dir ds41
hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf --local-dir ds41
cd ds41
ln -s DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-00006-of-00006.gguf

llama-server -m DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-IQ2_XXS-00001-of-00006.gguf -ngl 999 \
  --reasoning on --reasoning-effort 50

The Engram table is memory-mapped and read one row per token, so it needs neither VRAM nor RAM. To keep the whole
backbone on GPU, you need about 170 GB of VRAM; otherwise, offload experts with --n-cpu-moe.

Quality, limitations and responsible use

Benchmarks, refusal measurements and limitations are those of the source model; see
AtomicChat's model card. This
model answers requests the original declines. You are responsible for how you use it and for following the laws that
apply to you.

Credits and license

• Base model: deepseek-ai/DeepSeek-V4.1-Flash-1017.
• GSQ-RCO quantization, Engram file and converter: ISTA-DASLab.
• Uncensored edit: AtomicChat.

MIT, as the base model and both source repositories.