InfinimindCreations/Spark-X2.5-4B-uncensored-GGUF

Spark-X2.5-4B Uncensored, GGUF

Quantized GGUF builds of InfinimindCreations/Spark-X2.5-4B-uncensored, an abliterated Spark-X2.5-4B with 0 refusals across 400 harmful prompts and no measurable capability loss (perplexity change -0.06 %). Evaluation details and method are on the source model card.

File	Size	Notes
Spark-X2.5-4B-uncensored-Q4_K_M.gguf	2.60 GB	recommended default
Spark-X2.5-4B-uncensored-Q6_K.gguf	3.38 GB	higher quality
Spark-X2.5-4B-uncensored-Q8_0.gguf	4.38 GB	near-lossless
Spark-X2.5-4B-uncensored-BF16.gguf	8.23 GB	unquantized reference

Two things that will cost you time if you skip them

1. Stock llama.cpp and stock Ollama cannot load this

Spark-X2.5 support lives in a fork, XHToken/llama.cpp. Upstream llama.cpp has no Spark2_5 converter and Ollama 0.33.3 rejects the file with failed to validate GGUF with llama-quantize without compatibility patches. Measured, not assumed: we hit exactly that error.

2. The answer can come back empty

Spark writes into a reasoning block by default. Through an OpenAI-compatible endpoint the visible content is then empty while the actual answer sits in reasoning_content. Pass enable_thinking: false, or run Ollama with --think=false.

curl -s http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "messages": [{"role": "user", "content": "..."}],
  "chat_template_kwargs": {"enable_thinking": false}
}'

llama.cpp

git clone https://github.com/XHToken/llama.cpp.git llama.cpp-spark
cd llama.cpp-spark
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
cmake --build build --target llama-server -j 16

./build/bin/llama-server -m Spark-X2.5-4B-uncensored-Q4_K_M.gguf -ngl 99 -c 8192

Note: llama-cli may drop into interactive mode even with -no-cnv. Use llama-server for scripted checks.

Ollama

Ollama has to be rebuilt against the same fork:

git clone https://github.com/XHToken/llama.cpp.git llama.cpp-spark
git clone https://github.com/ollama/ollama.git ollama-spark
cd ollama-spark
export OLLAMA_LLAMA_CPP_SOURCE="$(cd ../llama.cpp-spark && pwd)"
cmake -S . -B build && cmake --build build --parallel 8

./ollama serve            # first terminal
./ollama create spark-uncensored -f ./Modelfile      # second terminal
./ollama run spark-uncensored --think=false

A ready Modelfile is included in this repository.

Verification

Each quantization was produced with llama-quantize from the fork, off the same BF16 export. The Q4_K_M build was loaded on GPU and checked twice: a raw completion ("The capital of France is" gives " Paris.") and a chat request that the base model refuses, which this build answers.

Disclaimer

Refusal behavior has been removed. This model will answer requests the base model declines. Published for research on alignment, refusal mechanisms and evaluation methodology. You are responsible for your use of it and for compliance with applicable law.