Best Local Text-to-Song AI Models (2026): YuE2, MiniMax Music 3, ACE-Step 1.5 XL and HeartMuLa

Want to make full songs with vocals and lyrics on your own GPU, without paying for Suno or uploading anything? In 2026 you finally can. Three open models released in the last few months, YuE2-3B (September), MiniMax Music 3 (August) and ACE-Step 1.5 XL (April), now run natively in ComfyUI and get close to commercial song generators. In this guide I compare every local text-to-song model worth trying in November 2026: how good each one is, how much VRAM it needs, the exact files and ComfyUI folders, how to write lyrics for it, and, most importantly, which licenses let you use the songs commercially.

Quick Comparison Table

These are the open-weight models that generate complete songs (vocals plus accompaniment) from lyrics and a style description. Dates are when the version I list was released. VRAM figures come from the model authors unless I mark them as an estimate.

What Changed in the Last Three Months

Before writing this I went through the ComfyUI changelog, the Hugging Face model pages and the project repos for everything released between August and early November 2026. Local song generation moved faster in these three months than in the whole year before:

• Aug 13, 2026: MiniMax released MiniMax Music 3 as open weights, with day-0 ComfyUI support (v0.33.1) and new Music3 Text Encode and Empty Music3 Latent Audio nodes. Community GGUF quants of its DiT appeared within days.
• Sep 11-15, 2026: MAP released YuE2-3B, and ComfyUI v0.36.0 added it natively with nodes YuE2 Generate ABC, YuE2 Generate Music and Empty YuE2 Latent Audio. Max song length in ComfyUI went up to 900 seconds. v0.37.0 added CFG control, and v0.38.0 made YuE2’s autoregressive stage faster.
• Sep 16, 2026: HeartMuLa released MuLaCover, a cover and remix model that keeps a song’s melody while you change genre and lyrics.
• Sep 29, 2026: ComfyUI v0.38.0 gave ACE-Step 1.5 CUDA graphs and the memory compiler, and removed the torchaudio dependency that used to break audio installs.
• Nov 1, 2026: Plenio 1.5, a ComfyUI song studio that uses a local LLM to write lyrics and then YuE2 or MiniMax Music 3 to perform them.

Older models such as YuE v1 (7B), DiffRhythm v1, SongBloom and MusicGen still work, but for songs with vocals they are now clearly behind the four models below.

1. YuE2-3B: Best Quality With an Editable Score

YuE2 is the follow-up to YuE from the Multimodal Art Projection (MAP) team. It is a 3B model that works in three stages: an autoregressive model first writes a symbolic score of the melody and chords in ABC notation, a flow-matching model turns that plan into acoustic latents, and a VAE renders 48 kHz stereo audio. Because the score is plain text, you can edit the melody before the audio is rendered, which no other open model offers.

On MAP’s WildSongBench (192 prompts), YuE2 with best-of-8 sampling scored 6.9632 against 6.8721 for Suno v5; single-sample YuE2 scored 6.7316. These are the authors’ numbers on their own benchmark, so treat them as a sign that it is in the same league rather than proof it beats Suno. Independent testers report convincing arrangements but vocals that still sound a little synthetic. The model card lists English and Chinese; testers also got usable results in French, Portuguese, German and several other languages.

Hardware

MAP measured about 11 GiB peak VRAM on an RTX 4090 (up to 14 GiB at maximum context) and generated a 3.6-minute song in 71 seconds. They recommend 24 GB of free system RAM. In ComfyUI, the INT8 checkpoint roughly halves the weight size, which makes 12-16 GB cards workable (my estimate based on file size).

YuE2 in ComfyUI: Files and Folders

1. Update ComfyUI to v0.36.0 or newer (v0.38.0+ for the faster sampler).
2. Download one checkpoint from Comfy-Org/YuE2 and put it in ComfyUI/models/checkpoints/:
   • ComfyUI/models/checkpoints/yue2_3b_bf16.safetensors (7.8 GB, 24 GB cards)
   • ComfyUI/models/checkpoints/yue2_3b_int8_convrot.safetensors (3.96 GB, 12-16 GB cards)
3. For cover and reference-audio templates, ComfyUI also uses the SheetSage2 transcriber: ComfyUI/models/audio_encoders/sheetsage2_bf16.safetensors.
4. Open Templates > Audio and pick a YuE2 template. Enter a style prompt and lyrics, set the duration, and queue.

Outside ComfyUI, the official route is the Python wheel on Linux:

python -m pip install huggingface-hub==0.36.2
hf download m-a-p/YuE2-3B yue2_infer-0.1.5-py3-none-any.whl --local-dir .
python -m pip install ./yue2_infer-0.1.5-py3-none-any.whl

Then load it with YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B", device="cuda"). Planning modes are cot="full", "melody" and "off".

Catch: the weights are CC BY-NC 4.0. That makes YuE2 great for demos, personal projects and research, but not for monetised YouTube music or client work.

2. MiniMax Music 3: Five-Minute Songs, Structured Captions

MiniMax Music 3 is the largest open song model right now. It is a hierarchical system: an 8B “global” LLM (initialised from Qwen3.5-8B) plans the song structure, a 0.6B local LLM adds acoustic detail, and a 2.4B flow-matching module plus a 123M Flow-VAE render 32 kHz, 16-bit stereo audio. It generates complete songs of up to about five minutes in one pass, and its arrangements and song structure are the most “radio-like” of the open models.

MiniMax Music 3 in ComfyUI: Files and Folders

Update ComfyUI (v0.33.1 or newer, v0.37.0+ fixes noisy output with CUDA graphs), then open Template Library > Audio > MiniMax Music 3 Text to Music. The template offers to download these files from Comfy-Org/MiniMax-Music-3:

For 8-12 GB cards, community GGUF quants of the DiT are on Abiray/MiniMax-Music3-GGUF: Q8_0 is 2.70 GB, Q5_K_M 1.79 GB and Q4_K_M 1.49 GB. Put the .gguf in ComfyUI/models/diffusion_models/ and load it with Unet Loader (GGUF) from the ComfyUI-GGUF pack (see my ComfyUI-GGUF guide for install steps). Keep the official text encoder and VAE from the Comfy-Org repo.

The workflow has four controls that matter: caption, lyrics, max_duration (the template defaults to 60 seconds) and tiled_decode, which lowers VRAM at a small risk of seams. Turn tiled decode off if you have 24 GB or more. Output lands in ComfyUI/output/audio/.

The Three-Part Caption

MiniMax Music 3 expects a structured caption with three sections: Global Metadata (genre, tempo, key, mood), Vocal Details (voice type, gender, delivery) and Arrangement (instruments and how they enter). Lyrics use square-bracket tags such as [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Solo], [Instrumental] and [Outro]. Tags set the structure; the words set the mood.

License: the MiniMax-Music3 Community License lets companies with under $20 million yearly revenue use it commercially. Bigger companies need a commercial license, which Comfy resells.

3. ACE-Step 1.5 and 1.5 XL: Fastest, MIT Licensed

ACE-Step 1.5 (January 28, 2026) is the model I recommend most people start with. It pairs a diffusion transformer (DiT) with an optional Qwen3-based “5 Hz” language model that plans the song. It is extremely fast: Comfy reports a full song in under 2 seconds on an A100 and under 10 seconds on an RTX 3090. It supports 50+ languages, tracks from 10 seconds to 10 minutes, batches of up to 8 songs, covers, repainting part of a song, and LoRA training from as few as 8 songs.

On April 2, 2026 the team added ACE-Step 1.5 XL, a 4B DiT with noticeably better vocals and mixing. Both sizes come in three flavours:

The official VRAM guide is very practical: 6 GB or less, use the 2B turbo with no LM; 6-8 GB, add the 0.6B LM; 8-16 GB, 2B turbo or sft with the 0.6B or 1.7B LM; 16-20 GB, XL turbo with the 1.7B LM; 24 GB+, XL sft with the 4B LM. The XL Turbo model card says 12 GB works with CPU offload plus INT8, 16 GB with offload, and 20 GB without offload.

ACE-Step 1.5 in ComfyUI: Files and Folders

Easiest (all-in-one): download ace_step_1.5_turbo_aio.safetensors from Comfy-Org/ace_step_1.5_ComfyUI_files into ComfyUI/models/checkpoints/ and open the template ACE-Step 1.5 Music Generation AIO.

Split files (needed for XL):

• ComfyUI/models/diffusion_models/acestep_v1.5_turbo.safetensors or acestep_v1.5_xl_turbo_bf16.safetensors (about 18.8 GB in bf16)
• ComfyUI/models/text_encoders/qwen_0.6b_ace15.safetensors
• ComfyUI/models/text_encoders/qwen_1.7b_ace15.safetensors (or qwen_4b_ace15.safetensors for XL on 24 GB)
• ComfyUI/models/vae/ace_1.5_vae.safetensors

The split workflow uses UNETLoader, DualCLIPLoader (both Qwen encoders), VAELoader, TextEncodeAceStepAudio1.5 for tags and lyrics, EmptyAceStep1.5LatentAudio for the duration, KSampler and VAEDecodeAudio. Cover and Repaint are not in ComfyUI yet; use the official Gradio app for those:

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
uv run acestep        # web UI at http://localhost:7860
uv run acestep-api    # REST API at http://localhost:8001

ACE-Step 1.5 is MIT licensed, and the team says its training data is licensed, royalty-free or synthetic, so it is the safest choice if you want to publish or sell what you make.

4. HeartMuLa 3B: Apache 2.0, Many Languages

HeartMuLa is a music language model family that generates songs from lyrics and tags, with multilingual support the authors describe as covering “almost all languages”. The team recommends HeartMuLa-oss-3B-happy-new-year (Feb 13, 2026) with HeartCodec-oss-20260123, which outputs 48 kHz stereo. In January the repo and all weights moved to Apache 2.0. The authors say an internal 7B version is comparable with Suno, but only the 3B is public.

There is no official VRAM figure. The repo suggests --lazy_load true on a single GPU and, if you have two cards, splitting the parts: --mula_device cuda:0 --codec_device cuda:1. That second option is a nice fit for a dual-GPU rig like my 2x RTX 5080. For ComfyUI, use the community node benjiyaya/HeartMuLa_ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/benjiyaya/HeartMuLa_ComfyUI.git
pip install -r HeartMuLa_ComfyUI/requirements.txt

Follow that repo’s README for where it stores the HeartMuLa and HeartCodec weights, as it can change between versions. For covers and remixes, the new MuLaCover (Sep 16, 2026) takes reference audio or melody and chord MIDI and lets you change genre, instruments and lyrics.

5. LeVo 2, DiffRhythm 2 and Other Options

• LeVo 2 (Tencent SongGeneration v2): strong vocals and about 4B parameters, with 10, 22 and 28 GB VRAM tiers. Its custom license is limited to academic, research and education use. GitHub.
• DiffRhythm 2 (Oct 30, 2025, ASLP Lab): a fast diffusion song model under Apache 2.0. Good when you want speed and a permissive license, but behind ACE-Step 1.5 XL in quality. GitHub.
• Khala (May 2026, ~3.7B, CC BY-NC 4.0) and Muse (0.6B, Apache 2.0 weights) are research models worth watching.
• SongBloom needs a reference audio clip, and YuE v1 (7B) is replaced by YuE2.

Which Model Fits Your VRAM

System RAM matters too. Offloading a 4B DiT or MiniMax’s 8B planner needs 32 GB of RAM at minimum, and 64 GB is more comfortable. If ComfyUI drops to “Reconnecting” while loading a song model, it almost always ran out of RAM; see my ComfyUI Reconnecting fix guide.

How to Write Lyrics and Style Prompts

All four main models use the same basic idea: a style description (genre, mood, tempo, instruments, voice) plus lyrics split into tagged sections. A template that works well across models:

Style: city pop, upbeat, 112 BPM, groovy bass, bright synth pads,
warm female vocal, clean modern mix

[Intro]

[Verse]
Neon on the river, city in my chest
Every light is calling, I don't need to rest

[Chorus]
Turn it up tonight, we are never too late
Hold the moment tight, let the morning wait

[Bridge]
...

[Outro]

Tips that help with every model:

• Keep lines short and singable. 6-10 syllables per line and 4 lines per section give the cleanest vocals.
• Match duration to lyrics. If the lyrics are too long for the duration, models rush or skip lines. About 60-90 seconds per verse-plus-chorus is a good rule.
• Use [Instrumental] or [Solo] for breaks instead of leaving empty sections.
• Fix the seed when you like a take, then change only the lyrics or the style to iterate.
• For MiniMax Music 3, write the caption as the three labelled parts. For YuE2, you can edit the generated ABC score before rendering if the melody is close but not right.

Need a lyric writer? Any local LLM works. I use a model in Ollama; see my Ollama install guide if you have not set one up.

Plenio 1.5: An All-in-One Song Studio for ComfyUI

If you want a “Suno-style” experience inside ComfyUI, Plenio Music Production System 1.5 (Nov 1, 2026) chains everything: a local LLM writes the title, style and lyrics, then YuE2-3B or MiniMax Music 3 performs the song, followed by mastering. Version 1.5 adds singing your own lyrics exactly as written, 250 style templates, a creativity slider up to 200 and custom section tags such as [Choir]. Install it from ComfyUI Manager by searching “Plenio Music Production System”, or run comfy node install comfyui-plenio-music. It needs ComfyUI 0.37.0 or newer, and its “0 · System Check” template checks your GPU and model files. Repo: jplenio/Plenio-Music-Production-System.

Can You Use the Songs Commercially?

A permissive model license does not make every output safe. Do not prompt for a real artist’s voice or copy existing lyrics, and keep your prompts and seeds in case a platform’s content ID flags a track. This is not legal advice; check the license text for your situation.

Common Problems and Fixes

• Missing YuE2 or Music3 nodes: your ComfyUI is older than v0.36.0 / v0.33.1. Update, or switch to Nightly in the Desktop app. See ComfyUI errors and fixes.
• Noisy or static MiniMax output: fixed in v0.37.0; update ComfyUI.
• ACE-Step VAE decode crash on older GPUs: fixed in v0.37.0 for non-bf16 cards (RTX 20-series and older).
• Out of memory at the end of a long song: it is the VAE decode. Turn on tiled decode or shorten the song.
• Model file not showing in the loader: check the folder again. My ComfyUI model folder cheat sheet lists every folder, then press R to refresh.

Frequently Asked Questions

What is the best free local alternative to Suno in 2026?

For quality, YuE2-3B and MiniMax Music 3. For commercial use and low VRAM, ACE-Step 1.5 (or 1.5 XL on 16 GB+). All three run natively in ComfyUI.

Can I generate songs with vocals on an 8 GB GPU?

Yes. ACE-Step 1.5 turbo with the 0.6B LM is built for 6-8 GB, and it runs DiT-only under 4 GB.

Which local song model supports the most languages?

ACE-Step 1.5 lists 50+ languages and HeartMuLa says it covers almost all languages. YuE2 officially lists English and Chinese.

How long can local AI songs be?

ACE-Step 1.5 goes up to 10 minutes, MiniMax Music 3 about 5 minutes, and ComfyUI allows YuE2 up to 900 seconds, although its demos stop at 5 minutes.

Can I make instrumental music with these models?

Yes, with instrumental tags or empty lyrics, but dedicated music models are often better for background tracks. See my background music generator guide.

Can I use YuE2 songs on monetised YouTube?

No. YuE2 weights are CC BY-NC 4.0 (non-commercial). Use ACE-Step 1.5, HeartMuLa or MiniMax Music 3 (under its revenue limit) instead.

Final Thoughts

Local song generation is no longer a toy. If you want one model to install today, make it ACE-Step 1.5: it is fast, MIT licensed and fits almost any GPU. When you want the best possible song and the license fits your use, try YuE2 and MiniMax Music 3 side by side with the same lyrics.

This guide is part of a three-part audio series:

• Best Local AI Background Music Generators: royalty-free instrumentals and loops for videos, games and streams.
• Best Local Text-to-SFX AI Models: sound effects and foley from text or video.

For voices, see Fish Audio Drama 3 locally in ComfyUI and the Voicebox tutorial. To put your new song under a video, my ffmpeg guide to adding music without re-encoding shows how, and for the video itself see the best local AI video models.

Sources

• m-a-p/YuE2-3B model card (benchmarks, VRAM, license)
• ComfyUI Wiki: YuE2-3B native support
• ComfyUI changelog (v0.33.1 to v0.42.4)
• ComfyUI docs: MiniMax Music 3
• MiniMax: Music 3.0 announcement
• MiniMax-Music3 GGUF quants
• ACE-Step 1.5 GitHub (models, VRAM tiers)
• ACE-Step 1.5 XL Turbo model card
• ComfyUI docs: ACE-Step 1.5
• HeartMuLa heartlib GitHub
• IT-JIM: open-source music generator comparison (LeVo 2, Khala, Muse)
• ComfyUI Wiki: Plenio 1.5
• MindStudio: YuE2 hands-on