GGUF quantization of my merge qwen3.5-9b coding model DarkKitsune/qwen3.5-9b-qworus to Q5_K_M and Q5_K_L with importance matrices.

Q5_K_M: The best option for 8 GB GPUs. Configured with BF16 for ssm_alpha and ssm_beta weights, Q8_0 for ssm_out weights, and the default llama.cpp Q5_K_M config for all other weights. This provides high precision for coding agentic work while having a smaller memory footprint.

Q5_K_L: The best option for 12+ GB GPUs. Configured with Q8_0 for token embedding and output weights, BF16 for ssm_alpha and ssm_beta weights, Q8_0 for ssm_out weights, and the default llama.cpp Q5_K_M config for all other weights. This trades a higher memory footprint for higher precision in tensors that may be sensitive to quantization when working with very long contexts.

I recommend using unquantized KV cache or as low as Q8_0. Quantizing the KV cache too much has a detrimental effect on the model's ability to emit tool calls, especially with code.

This model is also decensored, due to the fact that both source models in the merge were already decensored.

60/40 DARE-TIES of empero-ai/Qwythos-9B-v2 and PeppX/Ornith-1.0-9B-Uncensored. This merge exists because I find Qwythos to be excellent for its genuine stability and its ability to bug fix based on stderr output, but Ornith is better in its ability to set up a well-organized project with scalability in mind. So my goal is to cleanly combine the abilities of both.

Both source models are decensored, so this merge is decensored by default as well. It may inherit Qwythos' 1M context support due to using it as the base model, but I don't have the resources to actually test 1M tokens. The MTP head is also most likely not retained.

This model is for testing, and not intended to be used for any serious projects.

The following YAML configuration was used to produce this model: