AI is happening. The future is updating.higgs-audio-v3-tts-4b-transformers
A `trust_remote_code` packaging of `bosonai/higgs-audio-v3-tts-4b` that loads with plain π€ transformers (no SGLang). The weights are the original checkpoint, copied unchanged; only a small `modeling_*.py` / `configuration_*.py` pair and an `auto_map` were added. The model is a standard **Qwen3-4B** backbone plus a fused multi-codebook audio embedding/head. Reference-audio encoding and waveform decoding use the transformers-native `bosonai/higgs-audio-v2-tokenizer` (`higgs_audio_v2_tokenizer`), loaded automatically on first use. - Generation uses Higgs' delay pattern across 8 codebooks (vocab 1026, incl. BOC/EOC specials); de-delay + decode are handled internally. - The codec runs in fp32 (decode is unstable in bf16); the LM backbone runs in the dtype you load it in (bf16 recommended). - License: research/non-commercial, inherited from the upstream checkpoint β see `LICENSE`.
A `trust_remote_code` packaging of `bosonai/higgs-audio-v3-tts-4b` that loads with plain π€ transformers (no SGLang). The weights are the original checkpoint, copied unchanged; only a small `modeling_*.py` / `configuration_*.py` pair and an `auto_map` were added. The model is a standard **Qwen3-4B** backbone plus a fused multi-codebook audio embedding/head. Reference-audio encoding and waveform decoding use the transformers-native `bosonai/higgs-audio-v2-tokenizer` (`higgs_audio_v2_tokenizer`), loaded automatically on first use. - Generation uses Higgs' delay pattern across 8 codebooks (vocab 1026, incl. BOC/EOC specials); de-delay + decode are handled internally. - The codec runs in fp32 (decode is unstable in bf16); the LM backbone runs in the dtype you load it in (bf16 recommended). - License: research/non-commercial, inherited from the upstream checkpoint β see `LICENSE`.