IndexTTS-2.5 — MLX (fp16)
MLX conversion of IndexTeam/IndexTTS-2.5
(revision d0aa86e75bb6f3437f3831e95056fa72842d89ef) for Apple Silicon, re-hosted for the
MLXEngine fleet. Consumed by
mlx-indextts2-swift (Swift-MLX port,
Apache-2.0) and by vanch007/mlx-indextts2
(Python-MLX, MIT), whose converter (convert_v25.py, rev 4a32c967) produced these tensors
with strict tensor-coverage verification (conversion_report.json).
Files
| File | Component | dtype |
|---|---|---|
gpt.safetensors | UnifiedVoice 2.5 GPT (24×1280, CAMPPlus speaker projection, language embedding, emotion conditioner) | fp16 |
codec.safetensors | EnhancedCodec (Vocos/ConvNeXt encoder+decoder, 8192×8 FVQ) | fp16 |
s2mel.safetensors | S2Mel — length regulator + DiT/CFM | fp16 |
bigvgan.safetensors | BigVGAN v2 22 kHz 80-band 256× vocoder (NVIDIA, MIT) | fp16 |
multilingual_zh_ja_yue_char_del.tiktoken | tiktoken BPE vocabulary (60 509 tokens incl. specials) | — |
model.safetensors | facebook/w2v-bert-2.0 (MIT) — the semantic front-end, unchanged | fp32 |
config.yaml / config.json / model_manifest.json / conversion_report.json | checkpoint config + conversion provenance | — |
feat1.pt / feat2.pt / wav2vec2bert_stats.pt | emotion matrices + w2v-BERT feature stats (torch pickles, as upstream) | — |
License
The model weights (gpt, codec, s2mel, the tokenizer and the feat/stats files) are
licensed under the bilibili Model Use License Agreement — see LICENSE (verbatim copy).
Commercial use is permitted; per §2.2 an entity whose products exceeded 100 million monthly
active users in the preceding month or RMB 1 billion annual revenue in the preceding year
must request a separate license from bilibili. Downstream recipients are bound by the same
Agreement (§3.4(a)); retain the notices and this LICENSE with every copy (§3.4(b)).
bigvgan.safetensors is NVIDIA's BigVGAN v2 (MIT). model.safetensors is Meta's
w2v-bert-2.0 (MIT).
Derivative-work statement (§4.1(a)): Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work. (The modification here is a dtype conversion + tensor re-layout for MLX.)