HeartMuLa-oss-3B / HeartCodec-oss — MLX conversion (mlx-heartmula-swift, AB-T-0376)

Licence: Apache License 2.0 (text in LICENSE, from HeartMuLa/heartlib @ 5858ca8).

Grant chain (read 2026-10-28, AB-L-0307):
  • HeartMuLa/heartlib LICENSE is the verbatim Apache-2.0 text (added 2026-01-19); its README News: "20 Jan. 2026 — License
    update: We update the license of this repo and all related model weights to Apache 2.0."
  • HeartMuLa/HeartMuLa-oss-3B-happy-new-year @ 41f6fc6, HeartMuLa/HeartCodec-oss-20260123 @ f889dab and HeartMuLa/HeartMuLaGen @ 2e18e01 carry `license: apache-2.0` on their model cards and
    NO LICENSE file in the repository; the grant is the card statement plus the heartlib README above.
  • tokenizer.json is the Llama-3.2 BPE vocabulary (128 256 ids). HeartMuLa's weights are MuLa Labs' own (the backbone uses
    the Llama-3.2 architecture through torchtune builders; no Meta weights are loaded). The worst reading of the vocabulary
    file is the Llama 3.2 Community License, which permits commercial use.
  • Not on this path: MuQ-MuLan (CC-BY-NC-4.0; the pipeline passes muq_mulan=None and feeds a zero vector) and MuLaCover.

Conversion: decode path only (SQ-Codec encoder, RVQ EMA buffers, vq_embed.project_in dropped); weight_norm folded in fp32 with torch._weight_norm(v, g, 0); Conv1d (O,I,K)->(O,K,I); ConvTranspose1d (I,O,K)->(O,K,I); RVQ codebooks (1,8192,32)->(8192,32); SQ decoder list renamed; then .to(float16)

Redistribution (Apache License 2.0, section 4):
  • This repository is a modified form of HeartMuLa/HeartCodec-oss-20260123 @ f889dab0532cfa4bf459f2a3367eb6d346b8eeda, by the HeartMuLa team, under the Apache License 2.0 (LICENSE).
  • What was changed: the decode path only (the SQ-Codec encoder, RVQ EMA buffers and vq_embed.project_in dropped); weight_norm folded in float32; Conv1d and ConvTranspose1d weights transposed to MLX layout; RVQ codebooks reshaped; the SQ decoder module list renamed; then every tensor cast to float16.
  • Converted by xocialize for mlx-heartmula-swift (https://github.com/xocialize/mlx-heartmula-swift, MIT). Not endorsed by the
    HeartMuLa team or MuLa Labs.