encoder.mlpackage: FP16 speech encoder
projection.mlpackage: FP16 projection from encoder space to Qwen embedding space
canary_decoder_stateful.mlpackage: FP16 stateful autoregressive decoder with KV cache
canary_lm_head.mlpackage: FP16 LM head that maps decoder hidden states to logits

Base Model

Original model: nvidia/canary-qwen-2.5b
Original license: CC-BY-4.0
Original architecture: FastConformer encoder + Qwen decoder with projection and LoRA adaptation

Please review and comply with the original model card and license terms:

Included Artifacts

File	Precision	Purpose
`encoder.mlpackage`	FP16	Speech encoder
`projection.mlpackage`	FP16	Encoder-to-LLM projection
`canary_decoder_stateful.mlpackage`	FP16	Stateful decoder with KV cache
`canary_lm_head.mlpackage`	FP16	Decoder hidden states to logits

This repo contains model artifacts only.
The decoder is separated from the LM head.
The stateful decoder is intended for macOS 15 / iOS 18 era CoreML state support.
Long-audio chunking, prompt formatting, and transcript stitching live in the runtime layer and are not included here.

This conversion is based on: