Add model card with performance benchmarks

Upload custom weights (codec_embed, timespk_embed, codec_head, face_linear)

Upload MLX Qwen2-0.5B backbone (fp32)

initial commit