Current Model Catalog
Useizwi list (or GET /v1/models) to see the live, currently enabled catalog.
Those endpoints only show variants that are enabled for download/use.
Izwi accepts many legacy aliases (for example lowercase IDs), but the canonical IDs below match izwi list output.
Text-to-Speech (TTS)
Kokoro-82Mrequiresespeak-ng: macOS, Linux, Windows
Voxtral-4B-TTS-2603 includes bundled voice assets licensed under CC BY-NC
4.0 and supports 20 preset voices with 24 kHz output.
VibeVoice-1.5B is a Microsoft long-form TTS model with reference-voice
cloning. It uses saved or direct reference voices rather than built-in speaker
presets.
FishAudio-Drama-3is Fish Audio’s Drama 3 (Fish Audio Community License v1.0), a 35.9B-parameter mixture-of-experts TTS model with about 3B parameters active per frame: a Qwen3.5-MoE trunk and a 13-layer codebook tower write codes that the Fish S2 codec turns into 44.1 kHz audio. It reads the language from the text. Without a reference it speaks in a voice its samplingseedsets (the same seed gives the same voice); with reference audio and its transcript, or a saved voice that has one, it speaks in that voice. Tags in the text, such as[sighs],[laughs],[pause]or<whisper>…</whisper>, direct the delivery where they stand (see Text-to-Speech). Long text is spoken in chunks of about 300 bytes, each chunk’s prompt carrying the audio of the first and the latest chunks so the voice holds across the text, and audio streams as it is generated. The download is about 41.5 GB, and Izwi estimates about 40 GB of memory: the weights take about 37 GiB on CPU, Metal and CUDA (compute capability 8.0 or newer). A machine that cannot hold it is refused when the model loads, before any weights are read. The license is free for research and non-commercial use, allows commercial use only below USD 500,000 in annual revenue, and grants no right to offer the model to third parties as a hosted speech service, such as an API on Izwi’s server; that needs a license from Fish Audio.izwi pullprints these terms before it downloads, andGET /v1/admin/modelsreports them inlicenseandlicense_notice. Built with Fish Audio.
Chatterbox-500Mis Resemble AI’s English TTS model (MIT code and weights): a 0.5B Llama speech-token model with classifier-free guidance, then S3Gen flow matching and a HiFT vocoder, with 24 kHz output. It has one built-in voice (default) and clones from reference audio alone, without a transcript. Every utterance carries Resemble AI’s imperceptible PerTh watermark (the catalog labels it Watermarked), whose checkpoint is downloaded from GitHub rather than Hugging Face (see Manual Downloads).
Maya1-3Bis Maya Research’s English voice-design model (Apache-2.0): a Llama 3.2-3B decoder that writes SNAC audio codes, 24 kHz output. Its voice comes from a written description (instructions), such as “Realistic female voice in the 30s age with a british accent. Low pitch, warm timbre, slow pacing.” Nine described voices are built in (see Voice Presets). Twenty tags placed in the text (<laugh>,<sigh>,<whisper>, …) make it laugh, sigh or whisper there. At 3B parameters it is slower than real time on Apple Silicon: on a 16 GB M1 Pro with Metal one request takes about 4.5 times as long as the speech it generates (3 times withMaya1-3B-GGUF). It has not yet been measured on NVIDIA GPUs. The download (about 6.7 GB) includes the SNAC 24 kHz decoder from its own repository.Maya1-3B-GGUFis the same model with its decoder quantized to Q8_0 (llama.cpp’s conversion, published asmradermacher/maya1-GGUF): about 3.6 GB to download and half the weight memory, with the same voices and tags.
Maya2-Native-2.5Bis Maya Research’s Maya 2 Native (Apache-2.0), a TTS model for Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. A Qwen3-1.7B backbone predicts the first codebook of each 80 ms audio frame, a depth decoder fills in the other 31, and Kyutai’s Mimi codec turns the frames into 24 kHz audio. It speaks with one of 20 built-in voices (see Voice Presets), or clones a voice from reference audio and its exact transcript with SpeechBrain’s ECAPA speaker encoder; the clone can speak any of the 12 languages, not only the reference’s. A request names its language inlanguage(a code such ashi) or leaves it to be read from the text’s script. Tags in square brackets ([laughs],[sighs],[whispers], …) are performed on a best-effort basis. Numbers, dates, currency and similar are spoken as words in the request’s language. The download (about 5.5 GB) includes the Mimi codec (CC BY 4.0) and the ECAPA speaker encoder from their own repositories. Izwi estimates about 7 GB of memory on Metal and CUDA; on CPU it runs in F32, and its weights alone take about 10 GB. On a 16 GB M1 Pro with Metal one request takes about as long as the speech it generates (0.85 to 1 times), and four at once produce speech about 1.3 times faster than real time together; on that machine’s CPU a request takes 8 to 9 times as long as its speech. It has not yet been measured on NVIDIA GPUs.
Chatterbox-Multilingual-500Mis the same design trained on 23 languages, with its own speech-token model (t3_mtl23ls_v2) and grapheme tokenizer and the English model’s S3Gen, voice encoder, default voice and watermark. Izwi speaks 22 of its languages: Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish and Turkish. Chinese goes through upstream’s word segmenter (spacy-pkuseg) and Cangjie table, downloaded with the model. Japanese needs a kanji reader Izwi has not ported and returns400. A request names its language inlanguage(a code such asfror a name such asFrench); without one it speaks English. It uses upstream’s default (V2) speech-token model. In Izwi’s spot check (two sentences per language, transcribed by Whisper) it was less intelligible in Hebrew, Norwegian and Malay than in the other languages. Hebrew is spoken without vowel points and Russian without stress marks, as upstream does without its optional add-ons.
For built-in speaker IDs, see Voice Presets.LFM3-Audio-2B-GGUFis Liquid AI’s LFM3-Audio-2B (LFM Open License v1.0, free below US$10 million in annual revenue), one model that speaks, transcribes and holds voice conversations. It pairs a Q4_K_M language model with a Q8_0 mmproj file that holds the speech encoder, the MaskGIT depth decoder and the LiquidCodec decoder;izwi pull LFM3-Audio-2B-GGUFdownloads both (about 1.46 GB), and Izwi estimates about 3 GB of memory.LFM3-Audio-2B-Q8_0-GGUFkeeps the language model in Q8_0: about 2.20 GB to download and about 4 GB of memory. For speech it uses afemaleormalevoice in 19 locales (American English by default; see Voice Presets), streams 24 kHz audio, and splits long text into sentences; it has no voice cloning, voice description or speed control. It transcribes 18 languages, and in voice chat answers a spoken turn with interleaved text and speech, and can call tools, through the same speech-to-speech routes as LFM2.5 Audio. It is the default Compact voice model; LFM2.5 Audio stays selectable in Voice settings › Engine. It also streams realtime transcription in its native LiveASR mode. Prefix reuse is not supported yet. It runs on CPU, Metal and CUDA; CUDA is compile-checked only and has not yet run on an NVIDIA GPU.
Speech Recognition (ASR)
Nemotron-3.5-Transcribe-30B-A4B-FP8 is NVIDIA’s Nemotron 3.5 Transcribe
(OpenMDW-1.1): a speech encoder and the Nemotron-3-Diarization speaker
encoder feeding a 30B mixture-of-experts language model, with a CTC head
that times words. It transcribes a recording of up to 60 minutes in one
pass, in 40 languages (or detects the language), translates to or from
English, takes context terms, labels speakers and times words without a
forced aligner (see Transcription).
The download is 34.7 GB. It needs about 31 GiB of memory on Metal and
CUDA and about 33 GiB on CPU, plus about 48 MiB per request in flight;
machines that cannot hold it are refused when it loads, before any weights
are read. CUDA needs compute capability 8.0 or newer. It is offline only:
live transcription and voice mode do not use it. It can diarize on its own
(see Diarization), but it
cannot be the ASR model of a diarization pipeline.
Diarization and Alignment
Chat
Qwen3.8-27B-FP8 uses the separate Qwen38Chat architecture/model family.
Qwen3.5 compatibility or backend evidence is not applied to it implicitly.
The listed 28.74 GiB is checkpoint payload size, not an expanded runtime-memory
estimate. On CUDA, projections stay resident as
native block FP8 by default, with a
Q8_0 fallback.
Currently Disabled (Not Listed by izwi list)
These variants exist in the catalog but are not currently enabled for standard listing/download:
- Legacy Qwen3 chat IDs:
Qwen3-0.6B,Qwen3-0.6B-4bit,Qwen3-1.7B,Qwen3-1.7B-4bit Qwen3-14B-GGUFGemma-3-4b-it- TTS 8-bit and BF16 metadata variants such as
Qwen3-TTS-12Hz-0.6B-Base-8bitandQwen3-TTS-12Hz-1.7B-VoiceDesign-bf16.
Removed variants
The five mlx 4-bit Qwen3-TTS conversions (Qwen3-TTS-12Hz-0.6B-Base-4bit,
Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit, Qwen3-TTS-12Hz-1.7B-Base-4bit,
Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit, Qwen3-TTS-12Hz-1.7B-VoiceDesign-4bit)
are no longer supported: they never reach the end of speech and produce
unintelligible audio. Use the dense variant of the same name without -4bit.
The mlx 4-bit forced aligner (Qwen3-ForcedAligner-0.6B-4bit) is no longer
supported: it was dequantized at load, so it saved download size but no
memory. Use Qwen3-ForcedAligner-0.6B.
- Requests and
izwi pullnaming a removed ID return HTTP 400 naming its replacement. The replacement is not substituted automatically: the Qwen3-TTS replacements need more memory, and the aligner’s is a larger download. - Studio projects, queued jobs, transcription and diarization reruns, and the remembered model selection that name a removed ID use its replacement.
- A removed variant already downloaded stays on disk until you delete it. The
admin model list reports it under
removed_modelswith its size, andDELETE /v1/admin/models/<id>removes it.
Downloading Models
Via CLI
Via Web UI
- Open
http://localhost:8080 - Go to Models in the sidebar
- Click Download on a model
Managing Models
For the complete UI, CLI, and API workflow, see Model Management.View Downloaded Models
Get Model Information
Load a Model into Memory
Unload a Model
Delete a Model
Model Storage
Custom Model Directory
Manual Downloads
Some models may require manual Hugging Face access setup (for example gated repositories that need license acceptance):Model Status
Check status:
Quantization Notes
-4bit/-8bit/-bf16are reduced-precision variants.-GGUFvariants are quantized GGUF artifacts.- Smaller/quantized variants reduce memory and disk use at some quality/accuracy tradeoff.
izwi listshows enabled variants only. Some catalog metadata exists for experimental 8-bit/BF16 TTS artifacts; they are not downloadable.