Oswald-ASR-0.6-V2-GGUF

GGUF of jsbeaudry/Oswald-ASR-0.6-V2 for llama.cpp: the decoder (Q8_0, Q4_K_M) and the audio encoder as a multimodal projector (mmproj-…-Q8_0). Scores, data and recipe are on the model card.

Use this repo's own projector, not V1's: training adapted the audio encoder's attention, so the two encoders differ. Converting the -hf checkpoint needed two one-line changes to llama.cpp's conversion/qwen3vl.py (read audio_config at the top level when there is no thinker_config; map model.audio_tower.* like thinker.audio_tower.*); scripts/qwen3_asr/to_gguf.sh applies them.

Gold test sets WER 9.1 (V1 10.3); numbers recognised whole 90.3% (V1 70.7%).

llama-mtmd-cli -m Oswald-ASR-0.6-V2-Q4_K_M.gguf --mmproj mmproj-Oswald-ASR-0.6-V2-Q8_0.gguf --audio clip.wav -p ''