Skip to content

Models and licences

A model is what turns a prompt into an image, a video or a sound. The built-in models download the first time you use them, and you can add your own. Error? See Troubleshooting.

Which model to start with

  • Images: FLUX.2 Klein 4B (klein4b) is the lightest and the first-run pick on machines with less than 32 GB of memory. FLUX.2 Klein 9B (klein9b) is the pick from 32 GB of unified memory on a Mac or of GPU memory on a graphics card. On a 16 GB Mac, Klein and Z-Image Turbo need q8 or q4: see Memory.
  • Video: the Video tab starts on ltx-8, which comes closest to the full model. On a 16 GB Mac, or a card with 8 to 12 GB, start with LTX-2.3 4-bit (ltx-4): it needs the least memory of the LTX models. That advice comes from the app's memory estimate, not from a test on those machines. See Switch to the smaller model.
  • Audio: qwen3_tts reads text aloud and is the smallest download. yue2 makes songs and sa3 makes sounds.

All built-in models download without a Hugging Face account. Where the original repository needs a login, the alias points to an ungated mirror whose weights match the original.

Download size and memory

The Download column is the one-time size, kept in the Hugging Face cache (see Data folder). Where a size is missing, such as ltx-4 and ltx-8 on Windows and Linux, the Web UI's download dialog shows it, with the free space on the drive, before anything downloads. Memory is another figure: what a model needs while it runs depends on your machine and its quant. The Web UI's badges show what fits. See Memory.

Image models

Model Alias Download Steps Negative prompt License
FLUX.2 Klein 4B klein4b 16 GB 4 No Apache-2.0
FLUX.2 Klein 9B klein9b 35 GB 4 No FLUX Non-Commercial License v2.1
Z-Image Turbo zit 33 GB 10 Yes Apache-2.0
Krea 2 Turbo krea2 36 GB 8 No Krea 2 Community License Agreement
Ideogram 4 Instant ideo 28 GB (macOS), 24 GB (Windows, Linux) 8 No Ideogram Non-Commercial Model Agreement
MOSS-Image Turbo moss 28 GB 4 No Apache-2.0
MOSS-Image SFT moss-sft 27 GB 50 Yes Apache-2.0

The License column names each model's licence; Licences has the terms and links.

FLUX.2 Klein

klein4b and klein9b are distilled to 4 steps. They don't use a negative prompt. They also take up to four references: images of a person, object, place or style the result should keep.

Z-Image Turbo

zit uses the beta scheduler on every platform, so a prompt and seed look alike on macOS, Windows and Linux. --scheduler or the Web UI's Scheduler picks another.

Krea 2 Turbo

krea2 defaults to 8 steps and guidance 1.0, which turns classifier-free guidance off. Guidance above 1.0 turns it on, and each step then takes about twice as long. It doesn't use a negative prompt. Only Krea 2 Turbo and its fine-tunes are supported, not Krea 2 Raw. Like zit, it uses the beta scheduler by default on every platform, which draws more skin texture than the evenly spaced one.

The weights are large, about 26 GB for the transformer plus a 9 GB text encoder, so run it quantized on smaller machines:

Use q8 or q4. The text encoder isn't quantized, so it adds about 9 GB at any level.

q8 needs about 18 GB of system memory, and q4 fits 10–12 GB GPUs. Unquantized, Krea 2 needs about 35 GB of system memory and is very slow on machines with less.

q8 needs about 18 GB of system memory, and q4 fits 10–12 GB GPUs. Unquantized, Krea 2 needs about 35 GB of system memory and is very slow on machines with less.

Ideogram 4 Instant

ideo is Ideogram 4 Instant, fal's distilled Ideogram 4. It makes an image in 8 steps, and its guidance is built in, so a guidance setting is ignored.

  • Sizes must be 256–2048 pixels, in steps of 16. Presets beyond that, such as xl at 16:9, are refused before the model loads.
  • No start image, references or upscale while generating, and no negative prompt. The plain upscale of a finished image works.
  • LoRAs work. Ones made for Ideogram 4 load onto Instant, but they were trained on the base model, so check the results. On Windows and Linux a LoRA adds about 1 GB of GPU memory.

Captions. Ideogram 4 reads structured JSON captions. A plain prompt is wrapped into one for you, but such a sparse caption often comes out as a flat cut-out on a checkerboard background, especially for posters and product shots. For a fuller caption, let the prompt enhancer write one (Smart writes better ones than Fast), or write your own with --json-prompt or the Web UI's JSON mode.

The grey frame. Ideogram 4 sometimes returns a grey "Image blocked by safety filter" frame for harmless prompts. The app nudges the first denoising step slightly (a first-step sigma of 1.004), which makes this rarer without changing your prompt, seed or size. If a harmless prompt is still blocked, try First sigma 1.005 or 1.006. This is not a way around the filter: prompts it actually objects to are still blocked.

The first job builds the model in models/ideo in the data folder (~/.ziv by default). It downloads Instant's 18.6 GB transformer and converts it for your platform (FP8 on macOS, BF16 on Windows and Linux), takes the text encoder, VAE and tokenizer from an ungated mirror of the base model, and then deletes the download. An interrupted download resumes on the next run. The Web UI shows the downloads on the job card, then Build model.

Memory and speed

ideo takes q4 (NF4 on Windows and Linux) and q8 (FP8), saved once as ideo@q4 and ideo@q8. On macOS the unquantized model is the FP8 build.

  • macOS: at 1024×1024, unquantized and q8 peak at about 33 GB and q4 at about 25 GB, mostly briefly while the image is decoded at the end. On an Apple silicon Mac all three take about 15 s per step at that size; q4 is faster at smaller sizes.
  • Windows and Linux: the transformer streams to the GPU block by block. On a 10 GB GPU with 32 GB of system memory, q4 and q8 both take about 3.5 s per step at 1024×1024, about a minute per image (q8 peaks at about 17 GB of system memory, q4 at 14 GB). Unquantized it takes about 15 s per step and peaks at about 26 GB, so pick q4 or q8.

Every source is pinned to a checked revision, and the folder is rebuilt when a pin changes. The model is under the Ideogram Non-Commercial Model Agreement, which applies as soon as you download or use it: see LICENSE.md in the folder.

MOSS-Image

moss and moss-sft are MOSS-Image from OpenMOSS, an image model with a 5-billion-parameter transformer and a Qwen3.5-based text encoder, released under the Apache-2.0 license. It renders text well, in English and in Chinese, and follows long, detailed prompts.

Two models share one set of weights:

moss (Turbo) moss-sft
What it is MOSS-Image SFT with a community Turbo LoRA distilled from it MOSS-Image SFT as released
Steps 4 by default 50
Guidance Built in; a guidance setting is ignored with a warning 3.0
Negative prompt No Yes

Start with moss. Pick moss-sft when you want a negative prompt or the model as released.

moss takes about 30 s per image at 1024×1024 on a 32 GB Mac, quantized; moss-sft about 15 minutes at 50 steps. Both fit a 32 GB Mac at every level.

moss takes about 13 s per image at 1024×1024 on a 10 GB GPU, quantized; moss-sft about 2 minutes at 50 steps. Both run on a 10 GB card at every level; run them quantized, as unquantized Turbo takes about a minute per image.

moss takes about 13 s per image at 1024×1024 on a 10 GB GPU, quantized; moss-sft about 2 minutes at 50 steps. Both run on a 10 GB card at every level; run them quantized, as unquantized Turbo takes about a minute per image.

Memory, as peaks for Turbo images from 1024×1024 up to 4 megapixels, on a 10 GB GPU with 32 GB of system memory (Windows and Linux):

Level GPU memory System memory
q4 about 9 GB about 15 GB
q8 about 9 GB about 17 GB
None about 7 GB about 17 GB

For 4× creative upscales of a MOSS-Image picture on a 10 GB GPU, results up to about 8 megapixels fit (576×864 to 2304×3456 takes about a minute at q8); a 16-megapixel result runs out of memory, so use 2× creative or the plain 4× for a larger one.

  • Sizes are multiples of 16, from about 0.25 to 4 megapixels. The model is made for about 1–2 megapixels; above that, Turbo's 4 steps look softer, so use 6–8 steps.
  • Text. Put the words you want in the image in quotes in the prompt.
  • Start image, upscale, LoRAs and live previews work. LoRAs must be MOSS-Image LoRAs: see LoRAs. The Turbo LoRA stays on, and a LoRA of your own is added on top.
  • No JSON captions. MOSS-Image takes plain prompts, so --json-prompt and the enhancer's JSON format don't apply.

The first job builds the model in models/moss or models/moss-sft in the data folder from the shared download, which takes a moment. The community Turbo LoRA comes from akhaliq/MOSS-Image-SFT-4-steps, pinned to a checked revision.

Video models

Model Alias Download Needs Can do License
LTX-2.3 4-bit ltx-4 28.6 GB (macOS) The least memory of the LTX models. About 20 GB of system memory on Windows and Linux. Text or image to video, Extend, upscale LTX-2 Community License Agreement and Gemma Terms of Use
LTX-2.3 8-bit ltx-8 37.8 GB (macOS) Closer to the full model. About 27 GB of system memory on Windows and Linux; a Mac with 32 GB or more. The same LTX-2 Community License Agreement and Gemma Terms of Use
LTX-2.5 4-bit ltx25-4 71 GB (Windows, Linux), 28.8 GB (macOS) The newer LTX, with the memory need of ltx-4: about 21 GB of system memory on Windows and Linux. The same LTX-2.x Community License Agreement
MiniMax H3 h3 98 GB (Windows, Linux), 107 GB (macOS) Much larger and slower than LTX. About 18 GB of free system memory on Windows and Linux; a Mac with 32 GB. 4 or 8 steps. No upscale on macOS MiniMax H3 Community License Agreement

All of them make video with sound, from a prompt or a starting image (with an optional last frame), and can extend a clip. The first run of each downloads it and, on Windows and Linux, prepares it for your GPU, which takes a while. See Video.

On a 10 GB GPU with 32 GB of system memory, ltx-4 fits clips up to 20 s at m, 15 s at l and 10 s at xl, with or without upscaling. On a 32 GB Mac, the estimate was measured with ltx-4 for clips up to about 10 s at 704×448 and is extrapolated beyond that; Macs with 24 GB or less stream the transformer from disk, so the estimate is on the high side there. A clip's memory depends on the size and length of what is generated, so an extension costs the same however long its source is.

You can also convert an LTX-2.3 or LTX-2.5 fine-tune or a MiniMax H3 fine-tune into a model you pick by name.

On macOS, ltx-4 and ltx-8 download only the files the app loads, at fixed revisions, from ungated repositories, so no Hugging Face account is needed. Each figure includes the 8.1 GB Gemma 3 text encoder, which the two share, and nothing is prepared afterwards.

LTX-2.5

ltx25-4 is LTX-2.5, the newer LTX model, at 4 bits. It needs about as much memory as ltx-4: a PC with a 10 GB GPU and about 21 GB of system memory, or a Mac with 32 GB. It downloads without a Hugging Face account.

The download is about 28.8 GB and holds the model with its own Gemma 4 text encoder; nothing is prepared afterwards. A 5 s clip at s with Low memory is estimated at about 15 GiB.

The download is about 71 GB. The first run prepares a 4-bit transformer and an FP8 text encoder from it, about 22 GiB more on disk, and later runs reuse them. As a typical figure, a 5 s clip with sound at 1280×704, upscaled 2×, took about 200 s on a 10 GB NVIDIA card with 32 GB of system memory, the same as ltx-4 with that setup. It needs about 21 GB of system memory.

The download is about 71 GB. The first run prepares a 4-bit transformer and an FP8 text encoder from it, about 22 GiB more on disk, and later runs reuse them. As a typical figure, a 5 s clip with sound at 1280×704, upscaled 2×, took about 200 s on a 10 GB NVIDIA card with 32 GB of system memory, the same as ltx-4 with that setup. It needs about 21 GB of system memory.

  • Sampler. Text and image to video use LTX-2.5's own distilled sampler, which adds fresh noise at every step; a repeated seed still gives the same video. Extend uses the plain deterministic sampler, as ltx-4 does.
  • No 8-bit level for ltx25-4 itself. LTX-2.5 fine-tunes convert at q4 or q8 and run with LTX-2.5's settings.

MiniMax H3

h3 is MiniMax H3, a video model that is stronger on dynamic scenes than LTX and makes stereo sound with the picture. It is far larger, so it is slower and its download is bigger. It downloads without a Hugging Face account.

  • Steps: 4 or 8, with 8 the default. Four steps take about a quarter less time and give the same scene, slightly softer; fast motion can blur.
  • Length: clips are 17n+5 frames long (22, 39 … 124, 141 …), and a count in between moves up to the next. The default is 5 s (124 frames at 24 fps). The model was trained on 4 to 15 s clips; outside that range quality isn't guaranteed.
  • Sizes: the same presets as LTX, in multiples of 32 (64 with upscaling). xl is above the size the model was trained on.
  • Extend works from 24 fps clips whose size is a multiple of 32. See Extending a video.
  • LoRAs must be MiniMax H3 LoRAs. See LoRAs.
  • Some H3 checkpoints don't follow the start image. A few community checkpoints ignore the start image in image to video, first and last frame and Extend. If a converted model does, use stock h3 for those modes.

MiniMax H3 on macOS does everything above except upscale. It is built on a port to MLX. Image to video, first and last frame and Extend run in stages, with one component in memory at a time, and their figures were measured on an Apple silicon Mac with 32 GB of memory; the other figures are estimates.

  • Download and first use: about 107 GB. The first run converts the text encoder once into a 4-bit copy of about 15 GB (h3@mlx-te4).
  • Memory and time: a 5 s clip at s is estimated at about 15 GiB and 23 minutes with 8 steps, 14 minutes with 4. At l it is about 18 GiB and 77 minutes. The image modes need about the same as text to video: a 3 s image-to-video clip at s with 8 steps took about 10 minutes and peaked at about 14 GiB with no swapping, and is estimated at about 14 minutes. First and last frame and Extend peak at about the same, and a LoRA adds about 1.5 GiB. A Mac with 32 GB is the target.
  • Extend continues from the source's last frame alone, so the seam is weaker than on Windows and Linux.
  • No upscale or review. Image to video, first and last frame and Extend show a step counter and stop at a step boundary; text to video shows its phases but no step counter, and Skip and Quit take effect when the clip finishes (While a job runs).
  • A local folder given to -m can carry its own INT8 transformer in the MLX port's layout. A folder with a bf16 transformer is refused; convert a fine-tune instead.
  • Download and first use: about 98 GB. The first run builds a 4-bit copy of the transformer, about 12 GB, as h3@q4; it took about 4 minutes, and later runs reuse it. A fine-tune gets a copy of its own, named after it.
  • Memory: a 5 s clip at 864×480 uses about 6.3 GiB of GPU memory and about 17 GiB of system memory; plan for about 18 GB free. 1280×736 clips and 14 s clips at 864×480 fit a 10 GB GPU.
  • Upscale: a 2× upscale renders at half size, enlarges the result in latent space with H3's own upscaler (0.69 GB, downloaded on first use) and refines it at full size in 2 more steps. The clip keeps the sound of the half-size render and looks like a native full-size render.
  • Time: the CLI prints an estimate when a job starts, and the Workspace shows one under Duration (s), such as About 5 min (estimated). On a 10 GB NVIDIA card, a 5 s clip with sound:
Size (16:9) 8 steps 4 steps
s 768×448 about 4 min about 3 min
m 1024×576 about 6 min about 4 min
l 1280×704 about 9 min about 6 min

Longer clips take longer: 10 s at 864×480 took about 9 minutes, 14 s about 13. An upscaled l clip is estimated at about 5½ minutes. The estimate leaves out the one-time preparation.

  • Download and first use: about 98 GB. The first run builds a 4-bit copy of the transformer, about 12 GB, as h3@q4; it took about 4 minutes, and later runs reuse it. A fine-tune gets a copy of its own, named after it.
  • Memory: a 5 s clip at 864×480 uses about 6.3 GiB of GPU memory and about 17 GiB of system memory; plan for about 18 GB free. 1280×736 clips and 14 s clips at 864×480 fit a 10 GB GPU.
  • Upscale: a 2× upscale renders at half size, enlarges the result in latent space with H3's own upscaler (0.69 GB, downloaded on first use) and refines it at full size in 2 more steps. The clip keeps the sound of the half-size render and looks like a native full-size render.
  • Time: the CLI prints an estimate when a job starts, and the Workspace shows one under Duration (s), such as About 5 min (estimated). On a 10 GB NVIDIA card, a 5 s clip with sound:
Size (16:9) 8 steps 4 steps
s 768×448 about 4 min about 3 min
m 1024×576 about 6 min about 4 min
l 1280×704 about 9 min about 6 min

Longer clips take longer: 10 s at 864×480 took about 9 minutes, 14 s about 13. An upscaled l clip is estimated at about 5½ minutes. The estimate leaves out the one-time preparation.

MiniMax H3 is released under the MiniMax H3 Community License Agreement, published with the model on Hugging Face (MiniMaxAI/MiniMax-H3). It restricts use in some regions and above an annual revenue limit, requires commercial products to show "MiniMax H3" in their interface, and counts LoRAs as derivatives of the model. Read it before you use the model or its output; see Licences.

Audio models

Model Alias Makes Download License
YuE2 3B yue2 Songs and instrumentals about 8 GB CC BY-NC 4.0 with additional creator permission
Qwen3-TTS 1.7B qwen3_tts Speech about 4.5 GB, and about 4.5 GB more for cloning a voice and again for designing one, each the first time it is used Apache License 2.0
Stable Audio 3 Medium sa3 Ambience and sound effects, and instrumental music about 10 GB Stability AI Community License and Gemma Terms of Use

The alias is what you pass to ziv audio -m. What each kind does, and how long a take can be, is in Audio; what they need is under Audio memory. YuE2 and Stable Audio 3 Medium are restricted in some way: read their Licences before you use the output.

Audio memory

The audio models fit a 10 GB GPU on Windows and Linux, and a Mac with 32 GB of unified memory. Figures are peak GPU memory.

Model macOS, 32 GB Windows and Linux, 10 GB card Measured at
YuE2 3B (yue2) about 10.5 GiB about 6.7 GiB A take of the full 6:00
Qwen3-TTS 1.7B (qwen3_tts) about 12.0 GiB about 6.0 GiB A text read in parts, in a default voice or a recording
Stable Audio 3 Medium (sa3) about 6.1 GiB about 6.1 GiB A sound of the full 380 s

Designing a voice loads its own checkpoint and needs no more than the Qwen3-TTS row on Windows and Linux.

On Windows and Linux, system memory matters too: YuE2 peaks at about 11 GiB while it loads or runs, which leaves a computer with 16 GB of RAM little to spare. YuE2 keeps its weights in system memory and moves the part it needs to the GPU, which is how it fits a 10 GB card.

LoRA formats

A LoRA must be made for the model it is used with. What each model reads:

Models Formats Notes
Z-Image, FLUX.2 Klein, Krea 2 Standard LoRAs and LyCORIS LoKr, on every platform and quant level, and combinable Other LyCORIS formats are skipped with a warning
Ideogram 4 Ideogram 4 LoRAs, loaded onto Instant Trained on the base model, so check the results
MOSS-Image (moss, moss-sft) kohya, ai-toolkit and diffusers/PEFT, and LyCORIS LoKr Applied unmerged at any quant level; LyCORIS LoHa and full diff weights are skipped with a warning. On moss the Turbo LoRA stays on and yours is added on top
LTX, macOS LTX LoRAs as MLX loads them
LTX, Windows and Linux LTX LoRAs in a diffusers-compatible format An unsupported file or scale fails before generation with a message saying why
MiniMax H3 PEFT, ai-toolkit, kohya and ComfyUI formats A LoRA that changes the model's timestep embedding is refused, and on macOS so is one that changes its AdaLN projections or anything outside its transformer blocks and text refiner. A LoRA adds a few percent to the time, and can break the seam of an extension

LoRAs from other trainers, such as PEFT adapter names or Kohya names, are renamed to the names the model reads when they load; your file isn't changed. See LoRAs.

Licences

Every built-in model comes with terms. The Web UI shows them in the download dialog, beside the model picker and on the Models page, and the CLI prints them before a download. Downloading a model means you agree to its terms. The one-line summaries are a guide; the linked licence is what applies. The links lead to the model's original page; where that is gated, the app downloads from an ungated mirror, so you need no Hugging Face account.

Licence Supplier Covers Terms Links
FLUX Non-Commercial License v2.1 Black Forest Labs klein9b Non-commercial use of the model; images may be used commercially Model page
Krea 2 Community License Agreement Krea krea2 Free under $1M revenue, images included; you must filter harmful content Licence
Ideogram Non-Commercial Model Agreement Ideogram ideo Non-commercial, images included: no ads, client work or paid products Model page
LTX-2 Community License Agreement Lightricks ltx-4, ltx-8 Free under $10M revenue Licence
LTX-2.x Community License Agreement Lightricks ltx25-4 Free under $10M revenue Model page
MiniMax H3 Community License Agreement MiniMax h3 Not licensed in the EU, UK, South Korea or the USA; written permission needed over $20M revenue; commercial products show "MiniMax H3" Licence
CC BY-NC 4.0 with additional creator permission M-A-P yue2 Non-commercial; individual creators may monetise their outputs, companies need a licence Licence
Stability AI Community License Stability AI sa3 Free under $1M revenue; commercial use needs a free registration Model page
Gemma Terms of Use Google ltx-4, ltx-8, sa3 Free to use, also commercially; Gemma's prohibited use policy applies Model page
Apache License 2.0 Tongyi-MAI zit Free to use, also commercially Model page
Apache License 2.0 Black Forest Labs klein4b Free to use, also commercially Licence
Apache License 2.0 OpenMOSS moss, moss-sft Free to use, also commercially Licence
Apache License 2.0 Qwen The Fast and Smart prompt enhancer models Free to use, also commercially Licence
Apache License 2.0 Qwen qwen3_tts Free to use, also commercially Licence
Apache License 2.0 ByteDance Seed The SeedVR2 upscale models Free to use, also commercially Model page
Research-only training data BiRefNet and ViTMatte authors Background removal Non-commercial: ViTMatte was trained on Adobe's Deep Image Matting dataset, whose agreement allows models trained on it only for non-commercial use, and BiRefNet (MIT) partly on DIS5K, which is research-only Dataset terms

LTX-2.3 and Stable Audio 3 Medium come with Gemma's terms as well as their own, because a Gemma model is their text encoder.

ltx25-4 carries no Gemma terms: its text encoder is Gemma 4, which is under Apache 2.0, and so are the text encoder and upscaler of h3.

FLUX.2 Klein 9B. The licence limits the model and anything built from it to non-commercial use, but the images it makes may be used for any purpose, commercial ones included (section 2.d). They may not be used to train a model that competes with FLUX.

Stable Audio 3 Medium. Commercial use needs a free registration with Stability AI, and the licence ends once you or your affiliates earn more than $1M a year. Anyone who makes the model, or a product built on it, available to others must show "Powered by Stability AI" and pass on the notice "This Stability AI Model is licensed under the Stability AI Community License, Copyright © Stability AI Ltd. All Rights Reserved". Personal and research use needs none of this.

YuE2. Individual creators may publish and sell the songs they make. A company that uses the model commercially needs a licence from its authors.

Your own models

Converted fine-tunes, model folders, Hugging Face repositories and aliases all work as models. How to add each is in Use your own models; how a name is found is in the next section.

How a name is found

When you pass -m <name>, or pick a model in the Web UI, the app looks in this order:

  1. A local path, such as ./my-model or ~/models/my-model, or a Hugging Face repo, such as Tongyi-MAI/Z-Image-Turbo, is used as is. A repo downloads on first use.
  2. A folder models/<name>/ in the data folder is used if it exists.
  3. An alias, built in or your own.
  4. Otherwise the name is used as given, and the job stops if no model is found under it.

A folder in models/ wins over an alias of the same name. ziv-model list shows every model and alias.

For video, a name is accepted when it is ltx-4, ltx-8, ltx25-4 or h3, a converted LTX or MiniMax H3 model, a known LTX repo (dgrauet/ltx*, Lightricks/LTX-Video*, unsloth/LTX-2.5*) or MiniMax H3 repo (MiniMaxAI/MiniMax-H3*, unsloth/MiniMax-H3*), a folder laid out like a MiniMax H3 repository, or a local path containing ltx.

Built-in aliases

Alias Model Makes
klein4b FLUX.2 Klein 4B Images
klein9b FLUX.2 Klein 9B Images
zit Z-Image Turbo Images
krea2 Krea 2 Turbo Images
ideo Ideogram 4 Instant Images
moss MOSS-Image Turbo Images
moss-sft MOSS-Image SFT Images
ltx-4 LTX-2.3 4-bit Video
ltx-8 LTX-2.3 8-bit Video
ltx25-4 LTX-2.5 4-bit Video
h3 MiniMax H3 Video
yue2 YuE2 3B Music
qwen3_tts Qwen3-TTS 1.7B Speech
sa3 Stable Audio 3 Medium Sound and music

Sizes and memory needs of each are in their sections above, and their terms in Licences.

The Web UI shows a built-in model by its name and lists the alias under it on the Models page.