Model weights are not included in this repository.

Upstream model repos

Profile config source

Machine-readable profile metadata lives in:

configs/tested-profiles.json

Launch-script examples live in:

configs/*.sh

Main profiles

Profile Model file Context KV Reasoning/template
unsloth-65k-f16 Qwen3.6-27B-UD-Q6_K_XL.gguf 65,536 f16/f16 no-think
unsloth-65k-q8 Qwen3.6-27B-UD-Q6_K_XL.gguf 65,536 q8_0/q8_0 no-think
unsloth-128k-f16 Qwen3.6-27B-UD-Q6_K_XL.gguf 131,072 f16/f16 no-think
unsloth-128k-q8 Qwen3.6-27B-UD-Q6_K_XL.gguf 131,072 q8_0/q8_0 no-think
bartowski-128k-f16 Qwen_Qwen3.6-27B-Q6_K_L.gguf 131,072 f16/f16 no-think
bartowski-128k-q8 Qwen_Qwen3.6-27B-Q6_K_L.gguf 131,072 q8_0/q8_0 no-think

Notes from upstream docs

Qwen3.6 documentation describes thinking mode, non-thinking mode, and thinking preservation. It also recommends at least 128K context for preserving thinking capabilities in complex tasks. See the Qwen model card and docs linked above.

Reference profile cards

Reference-only machine-readable profile cards live under:

examples/profiles/qwen3.6-27b/

These are not eval definitions. They document the profiles used in the original local runs so others can compare their launch settings and result packages.