Quick Answer
MiniMax-Music3 (MiniMaxAI) is the top pick as of September 2026 because it is the newest eligible release; this ranking admits only text-to-audio models from labs with a published paper or leaderboard record and follows newest release first, then downloads, with no public benchmark yet for any candidate [4][5][7][9][1][3]. In order, the ranking is MiniMax-Music3 (MiniMaxAI), Magenta RealTime 2 (Google), Stable Audio 3 Optimized (Stability AI), Stable Audio 3 Medium Base (Stability AI), MusicGen Medium (Facebook), and MusicGen Small (Facebook) [4][5].
Key Takeaways
- Ranking note: scope admits only text-to-audio models from labs with a published paper or leaderboard record. All candidates have no public benchmark yet; ordering is newest release first, then downloads, with Facebook’s MusicGen entries admitted as established picks that bypass the recency gate. [1][2][3][4][5][6][7][8][9]
- MiniMax-Music3 by MiniMaxAI ranks first because its August release is the newest admitted release. [4][5][7][9] Its 2.4B parameters imply 1.2 GB estimated (params x 0.5 bytes) for 4-bit weights; that estimate does not establish total memory requirements or Apple Silicon performance. [4]
- Magenta RealTime 2 by Google ranks second, with a May 28 release and a CC-BY-4.0 license. [5] Its full-precision weights occupy 11.3 GB; a parameter-derived quantized-memory estimate is unavailable. [5]
- Stable Audio 3 Optimized by Stability AI precedes Stable Audio 3 Medium Base by Stability AI because their releases fall on May 18 and May 17, respectively; both use the stable-audio-community license. [7][9] Medium Base has 2.3B parameters: 1.15 GB estimated (params x 0.5 bytes) for 4-bit weights. [9]
- MusicGen Medium by Facebook is an established pick, ordered ahead of its same-day sibling by downloads: 1,949,426 versus 144,863 over the reported monthly window. [1][3] Its cc-by-nc-4.0 license carries a noncommercial restriction. [1]
- MusicGen Small by Facebook is an established pick with 591M parameters: 0.296 GB estimated (params x 0.5 bytes) for 4-bit weights. [3] Its cc-by-nc-4.0 license also restricts commercial use; the weight estimate alone does not establish which Mac can run it. [3]
How do local audio generation models compare on memory, context, licenses and benchmarks?
| Model | Org | Params | Quant/VRAM | Released (date) | License | Key benchmark (date) |
|---|---|---|---|---|---|---|
| MiniMax-Music3 [4] | MiniMaxAI [4] | 2.4B [4] | 4-bit weights: 1.2 GB estimated (params x 0.5 bytes) [4] | 2026-08-07 [4] | — | no public benchmark yet |
| Magenta RealTime 2 [5] | Google [5] | — | Full-precision weights: 11.3 GB [5] | 2026-05-28 [5] | CC-BY-4.0 [5] | no public benchmark yet |
| Stable Audio 3 Optimized [7] | Stability AI [7] | — | — | 2026-05-18 [7] | stable-audio-community [7] | no public benchmark yet |
| Stable Audio 3 Medium Base [9] | Stability AI [9] | 2.3B [9] | 4-bit weights: 1.15 GB estimated (params x 0.5 bytes) [9] | 2026-05-17 [9] | stable-audio-community [9] | no public benchmark yet |
| MusicGen Medium — established pick [1] | Facebook [1] | — | — | 2023-06-08 [1] | cc-by-nc-4.0 [1] | no public benchmark yet |
| MusicGen Small — established pick [3] | Facebook [3] | 591M [3] | 4-bit weights: 295.5 MB estimated (params x 0.5 bytes) [3] | 2023-06-08 [3] | cc-by-nc-4.0 [3] | no public benchmark yet |
Which audio generation models should you consider for an Apple Silicon Mac?
1. MiniMax-Music3
MiniMax-Music3 by MiniMaxAI ranks first because its August 7, 2026 release places it ahead under the release-date ordering.[4] Ranking note: newest release first, then downloads; no public benchmark yet. Scope: this ranking admits only text-to-audio models from labs with a published paper or leaderboard record, using Hugging Face’s text-to-audio pipeline tag. Established picks bypass the recency gate and follow the same ordering rule, without taking first place through that exemption.[1][3]
MiniMax-Music3 has 2.4 billion parameters, implying 1.2 GB of 4-bit weights, estimated (params x 0.5 bytes).[4] The repository lists 47.0 GB of F32 weights.[4] Those figures describe different things: a parameter-based weight estimate and the published weight files. Neither establishes the unified memory needed for local inference on an Apple Silicon Mac. Hardware planning needs a verified runtime and its total memory requirements; the weight estimate alone cannot justify a Mac memory recommendation.
Consider MiniMax-Music3 for exploratory local music-generation evaluation when trying a recent release matters more to your workflow than having a published benchmark comparison. The practical caveat is unverified deployment suitability: confirm Apple Silicon execution, usable quantized weights, supported generation length and license terms before committing to a workflow. Its position reflects release order, so the ranking does not establish audio quality or Mac performance.
2. Magenta RealTime 2
Magenta RealTime 2 by Google ranks second, with a Hugging Face publication date of May 28, 2026.[5] MiniMaxAI’s MiniMax-Music3 precedes it because its release is newer, dated August 7, 2026.[4] The ordering is newest release first, then downloads; Magenta has no public benchmark yet. The ranking admits only text-to-audio models from labs with a published paper or leaderboard record; Google’s accompanying paper is Live Music Models.[6]
Magenta RealTime 2 carries a CC-BY-4.0 license and lists full-precision weights totaling 11.3 GB.[5] Its Hugging Face repository recorded 8,759 downloads over the preceding 30 days as of September 25, 2026.[5] Download counts indicate uptake, but do not establish audio quality or execution speed. Consider Magenta for text-to-audio music experiments where its license suits the project; the available performance information does not establish a narrower use-case advantage.
For an Apple Silicon Mac, the hardware requirement remains unconfirmed. The listed weight size is not a complete unified-memory budget: runtime memory and a verified Mac execution path are unspecified. A parameter count is also unavailable, preventing a parameter-based quantized-weight estimate. The practical caveat is therefore local feasibility: verify runtime compatibility and measured memory consumption before choosing Mac hardware for this model.
3. Stable Audio 3 Optimized
Stable Audio 3 Optimized by Stability AI ranks third under the ordering rule, “newest release first, then downloads”: its May 18, 2026 publication follows the MiniMaxAI and Google entries.[4][5][7] The position reflects release timing, not demonstrated audio quality or Mac performance. Its benchmark status is “no public benchmark yet.” The model uses the stable-audio-community license and belongs to the family described in Stability AI’s “Stable Audio 3” paper.[7][8]
Hardware requirements remain unverified for Apple Silicon. A defensible memory estimate requires a verified parameter count, while a practical RAM recommendation also needs runtime overhead and a working Mac implementation. Neither a quantized weight footprint nor a tested Mac configuration can be established here. Context limits and supported output duration also remain unverified. Avoid choosing a Mac memory configuration on the strength of the “Optimized” name alone.
Consider this model for local text-to-audio evaluation if you can validate the runtime on your own machine. Start by checking installation support, generating a representative clip and measuring peak memory before committing to a workflow. The central caveat is that “Optimized” does not establish Apple Silicon acceleration or generation speed. Review the stable-audio-community license against your intended use before adopting the model.[7]
4. Stable Audio 3 Medium Base
Stable Audio 3 Medium Base by Stability AI ranks fourth under the release-date ordering, with a Hugging Face debut on May 17, 2026.[9] The ordering basis is newest release first, then downloads; the preceding releases come from MiniMaxAI, Google, and Stability AI.[4][5][7][9] The ranking admits only text-to-audio models from labs with a published paper or leaderboard record. Stable Audio 3 Medium Base has no public benchmark yet, so its position does not establish comparative audio quality.
The model has 2.3 billion parameters and lists 10.4 GB of F32 weights.[9] Quantized weight storage would be approximately 1.15 GB, estimated (params x 0.5 bytes) from the cited 2.3 billion parameters.[9] That calculation covers weights alone; it does not establish total unified-memory requirements or a working quantization path. Choosing an Apple Silicon Mac therefore requires validating the runtime and measuring memory use during generation. A specific Mac memory configuration cannot be recommended from the weight estimate alone.
Consider the model for exploratory local text-to-audio work where you can evaluate compatibility before integrating it into a workflow. Stability AI documents the family in its “Stable Audio 3” paper.[8] The license is stable-audio-community; review its terms against your intended use.[9] The practical caveat is that Apple Silicon performance and supported generation duration remain unverified here.
5. MusicGen Medium
MusicGen Medium by Facebook ranks fifth as an established pick, with a Hugging Face publication date of 2023-06-08 and 1,949,426 downloads over the reported 30-day window.[1] The ordering basis is newest release first, then downloads; download volume breaks the tie with Facebook’s MusicGen Small.[1][3] MusicGen Medium bypasses the recency gate as an established pick, rather than earning its position through a measured quality advantage.[1] Its benchmark status is no public benchmark yet.
The ranking admits only text-to-audio models from labs with a published paper or leaderboard record. Facebook documents the MusicGen family in Simple and Controllable Music Generation.[2] MusicGen Medium uses the CC-BY-NC-4.0 license.[1] Parameter count, quantized weight memory and supported context length are not established here. Consequently, a specific Apple Silicon Mac or unified-memory capacity cannot be recommended with confidence. Weight storage alone would also be insufficient to establish the memory needed during generation.
Consider MusicGen Medium for noncommercial music-generation experiments where the published family methodology provides a useful starting point.[1][2] Before committing hardware, verify that your chosen implementation supports Apple Silicon, then check its runtime memory requirements and generation limits. No verified Mac runtime or hardware-fit figure accompanies this entry. The practical caveat is licensing: the noncommercial restriction makes MusicGen Medium unsuitable as a default choice for a commercial workflow without separate permission.[1]
6. MusicGen Small
MusicGen Small by Facebook is an established pick that bypasses the release recency gate.[3] The ordering is newest release first, then downloads.[1][3][4][5][7][9] Its June 8, 2023 publication date and 144,863 monthly downloads place it sixth, behind Facebook’s MusicGen Medium, which shares that date and records 1,949,426 monthly downloads.[1][3] MusicGen Small has no public benchmark yet. The ranking admits only text-to-audio models from labs with a published paper or leaderboard record; Facebook’s accompanying paper is “Simple and Controllable Music Generation.”[2]
MusicGen Small has 591 million parameters and a listed F32 weight size of 2.4 GB.[3] Its estimated 4-bit weight payload is 295.5 MB—estimated (params x 0.5 bytes), using the cited 591 million parameters.[3] Treat that calculation as a weight-storage estimate, not a measured memory requirement or confirmation that a compatible quantized release exists. A specific Apple Silicon chip or unified-memory capacity cannot be recommended confidently without verified runtime support and total memory measurements.
Consider MusicGen Small for noncommercial music-generation experiments if you can establish a working local runtime on your Mac. Its license is cc-by-nc-4.0.[3] The practical caveat is that a modest calculated weight payload does not establish Apple Silicon compatibility, generation speed, or audio quality. Context limits and supported output duration are also unconfirmed, so check those requirements before committing to a local workflow.
How can you estimate audio model weight memory from parameter counts?
Estimate audio model weight memory by multiplying the published parameter count by 0.5 bytes per parameter for 4-bit weights, labeling the result “estimated (params x 0.5 bytes).” [3][4][9] Treat the calculation as a weight-storage estimate, not a measurement of memory use during generation.
MiniMaxAI’s MiniMax-Music3 has 2.4 billion parameters, giving approximately 1.2 GB for weights: estimated (params x 0.5 bytes). [4] The stabilityai/stable-audio-3-medium-base model from stabilityai has 2.3 billion parameters, giving approximately 1.15 GB: estimated (params x 0.5 bytes). [9]
The facebook/musicgen-small model from facebook has 591 million parameters, giving approximately 295.5 MB for weights: estimated (params x 0.5 bytes). [3] Such arithmetic does not establish that a compatible quantized checkpoint or local execution path is available.
Published checkpoint sizes answer a different question. MiniMaxAI lists F32 weights totaling 47.0 GB for MiniMax-Music3; that figure should remain separate from the parameter-based estimate. [4] Google’s google/magenta-realtime-2 lists full-precision weights totaling 11.3 GB, but no parameter count is specified here, so leave its parameter-based estimate unstated. [5]
For an Apple Silicon Mac, use these calculations to budget for weights. A weight estimate alone does not establish total runtime memory, supported audio duration, generation speed, or whether a model fits a particular Mac.
Which licenses apply to local audio generation models?
The licenses are cc-by-nc-4.0 for Facebook’s MusicGen Medium and MusicGen Small,[1][3] CC-BY-4.0 for Google’s Magenta RealTime 2,[5] and stable-audio-community for Stability AI’s Stable Audio 3 Optimized and Stable Audio 3 Medium Base.[7][9]
Facebook’s listed MusicGen models share a license identifier,[1][3] as do the listed Stability AI models.[7][9] Google’s model carries a different Creative Commons license identifier from Facebook’s models.[1][3][5] Keep those distinctions visible when comparing candidates; “open weights” is not a substitute for recording the actual license.
For an engineering evaluation, record the repository, model revision and applicable license together. Read the terms against your intended use, including commercial deployment, redistribution and modification. Check how the terms address generated audio separately from how they address model weights, rather than assuming the same conditions apply.
For MiniMaxAI’s MiniMax-Music3,[4] treat license verification as an outstanding step before adoption. A license is not confirmed in this comparison.
Choose a model whose documented terms match the project, then evaluate local execution. Download availability alone should not be your licensing decision.
What remains unverified about running these models on Apple Silicon Macs?
Apple Silicon compatibility, generation speed, peak unified-memory use and supported audio duration remain unverified for these candidates; the published metadata does not establish a tested Mac configuration.[1][3][4][5][7][9] A practical evaluation still needs a reproducible installation, a named execution backend and confirmation of which operations run on the GPU or fall back to the CPU.
Quantized weight arithmetic does not establish runtime memory requirements. MiniMaxAI’s MiniMax-Music3 lists 2.4B parameters: its weight-only footprint would be 1.2 GB at 4-bit, estimated (params x 0.5 bytes).[4] Such an estimate excludes runtime allocations and additional pipeline components. Quantized checkpoint availability, backend support and audio quality after quantization still need verification before using that estimate to choose a Mac.
Prompt limits, maximum generated duration and streaming behavior also remain unverified. A model name or downloadable checkpoint cannot establish whether generation works continuously, how quickly audio becomes available or whether repeated runs exhaust memory. Each candidate has “no public benchmark yet” under the comparison criteria, leaving no published basis here for ranking Mac execution speed or audio quality.
Licensing needs a separate check from execution. facebook’s musicgen-small and musicgen-medium list cc-by-nc-4.0.[1][3] MiniMaxAI’s MiniMax-Music3 licensing remains unverified in the available model metadata.[4] Confirm the applicable terms before incorporating generated audio into a commercial workflow.
Frequently Asked Questions
Which audio generation model should I evaluate first on an Apple Silicon Mac?
MiniMax-Music3 (MiniMaxAI) takes first place because its Hugging Face publication date, 2026-08-07, is the newest among the recent candidates. [4][5][7][9] Ranking note: newest release first, then downloads; no public benchmark scores any candidate. Eligibility admits only text to audio models from labs with a published paper or leaderboard record (Hugging Face pipeline tags: text-to-audio). Placement reflects release order; Apple Silicon performance remains unverified.
What is the ranking, and why are older MusicGen models included?
The order is MiniMax-Music3 (MiniMaxAI) [4], google/magenta-realtime-2 from google [5], stabilityai/stable-audio-3-optimized from stabilityai [7], stabilityai/stable-audio-3-medium-base from stabilityai [9], facebook/musicgen-medium from facebook [1], then facebook/musicgen-small from facebook. [3] Both facebook entries are established picks that bypass the 12-month recency gate. [1][3] Their shared publication date makes downloads the tiebreak: 1,949,426 for medium and 144,863 for small over the last 30 days as of 2026-09-26. [1][3]
How much memory would quantized weights use?
For 4-bit weight storage, MiniMax-Music3 (MiniMaxAI) has 2.4B parameters: 1.2 GB, estimated (params x 0.5 bytes). [4] stabilityai/stable-audio-3-medium-base has 2.3B parameters: 1.15 GB, estimated (params x 0.5 bytes). [9] facebook/musicgen-small has 591M parameters: 0.2955 GB, estimated (params x 0.5 bytes). [3] Those calculations describe weight storage only. Quantized checkpoint availability, compatible Mac runtimes and total runtime memory remain unverified.
Can I choose a model just from its download size?
Published weight sizes include 11.3 GB at full precision for google/magenta-realtime-2 [5], 10.4 GB in F32 for stabilityai/stable-audio-3-medium-base [9], and 2.4 GB in F32 for facebook/musicgen-small. [3] Download size alone does not establish a verified Mac memory requirement. Apple Silicon runtime compatibility, measured peak memory and generation speed remain unspecified, so a practical deployment choice still requires checking those details.
What licenses do these audio models use?
google/magenta-realtime-2 uses CC-BY-4.0. [5] stabilityai/stable-audio-3-optimized and stabilityai/stable-audio-3-medium-base use stable-audio-community. [7][9] facebook/musicgen-medium and facebook/musicgen-small use cc-by-nc-4.0. [1][3] MiniMax-Music3 (MiniMaxAI) has no license specified in this comparison. Check the applicable license terms before deployment, particularly for commercial use or redistribution. A position in this ranking does not establish permission for your intended use.
Are there benchmark scores or context limits I can compare?
Every candidate carries the status “no public benchmark yet”; no public benchmark scores these candidates against one another. Context limits and maximum generated audio duration are also unspecified. Published papers include Simple and Controllable Music Generation [2], Live Music Models [6], and Stable Audio 3. [8] Paper availability alone does not establish comparable quality scores, Apple Silicon speed or supported generation length.
Sources
- facebook/musicgen-medium model card (Hugging Face) — 2026-09-25
- Simple and Controllable Music Generation — 2023-06-08
- facebook/musicgen-small model card (Hugging Face) — 2026-09-25
- MiniMaxAI/MiniMax-Music3 model card (Hugging Face) — 2026-09-25
- google/magenta-realtime-2 model card (Hugging Face) — 2026-09-25
- Live Music Models — 2025-08-06
- stabilityai/stable-audio-3-optimized model card (Hugging Face) — 2026-09-25
- Stable Audio 3 — 2026-05-18
- stabilityai/stable-audio-3-medium-base model card (Hugging Face) — 2026-09-25