Muse and Llama (Meta): multimodal open weights with licence conditions

Contents

1. What it does for a company
2. Strengths and limits
3. Pricing
4. Questions to ask before choosing Meta’s models
5. Sources

Type	Open-weight model family
Pricing	Published
Website	huggingface.co

Meta no longer leads with Llama. Its developer site now lists the Muse family, and the newest Llama models, Llama 4 Maverick and Scout, date from April 2025⁠12Products and solutions for AI developersMetameta-llama/Llama-4-Maverick-17B-128E-InstructMeta (Hugging Face) · 5 Apr 2025.

Best for: Multimodal work across text, images, video and audio from one open-weight model.

Not the right fit for: Companies that want a licence without branding duties or revenue thresholds.

What it does for a company

Muse Spark Open, released on 15 October 2026, is a mixture-of-experts model with 816 billion parameters, 42 billion active per token, a one-million-token context and input in text, images, video and audio⁠3meta-models/Muse-Spark-Open-816B-A42BMeta (Hugging Face) · 15 Oct 2026. It needs at least 700 GB of GPU memory at NVFP4 precision. Fine-tuning is permitted, and Hugging Face’s TRL includes a LoRA example⁠34meta-models/Muse-Spark-Open-816B-A42BMeta (Hugging Face) · 15 Oct 2026TRL: Transformer Reinforcement LearningHugging Face. Vendor-reported, it scores 89.0 on GPQA Diamond⁠3meta-models/Muse-Spark-Open-816B-A42BMeta (Hugging Face) · 15 Oct 2026. For smaller deployments Meta also released Muse Glimmer 30B under Apache 2.0, which fits in 24 to 32 GB at about 4-bit precision⁠5meta-models/Muse-Glimmer-30BMeta (Hugging Face).

The licences are what set Meta apart. The Muse Model License requires a "Built with Muse Spark Open" notice, requires a separate agreement for operators offering the model as a service with more than $20 million in revenue, applies a usage policy and excludes users under 18⁠3meta-models/Muse-Spark-Open-816B-A42BMeta (Hugging Face) · 15 Oct 2026. The Llama 4 Community License, which still governs Llama 4, requires a licence request above 700 million monthly users, a "Built with Llama" notice and "Llama" at the start of any derivative model’s name, and follows an acceptable-use policy⁠6Llama 4 Community License AgreementMeta (Hugging Face) · 5 Apr 2025.

Strengths and limits

Strengths

• The broadest input range of the families reviewed: text, images, video and audio.
• Fine-tuning is permitted, with a LoRA example in Hugging Face’s TRL.
• Muse Glimmer 30B is Apache 2.0 and fits in 24 to 32 GB at about 4-bit.

Limits

• Muse Spark Open’s licence requires a "Built with Muse Spark Open" notice and a separate agreement for service operators above $20 million revenue; it also needs at least 700 GB of GPU memory.
• Llama 4’s licence requires a licence request above 700 million monthly users, a "Built with Llama" notice and "Llama" at the start of derivative model names.

Choose another option if: You want a plain licence (DeepSeek, Gemma 4, gpt-oss) or a fully open model (OLMo).

Pricing

Weights are free under Meta’s licence terms.

Questions to ask before choosing Meta’s models

• Can we meet the branding duties in our own products?
• Will our use stay below the revenue and user thresholds, now and in three years?
• Do we need audio and video input, or would a text-and-image model do?
• Would the Apache-licensed Muse Glimmer 30B meet our needs without the flagship’s terms?

See how Meta’s models compare with the other options in the main guide.

Sources

1. Products and solutions for AI developers. Meta. Accessed 9 Nov 2026.
2. meta-llama/Llama-4-Maverick-17B-128E-Instruct. Meta (Hugging Face), 5 Apr 2025. Accessed 9 Nov 2026.
3. meta-models/Muse-Spark-Open-816B-A42B. Meta (Hugging Face), 15 Oct 2026. Accessed 9 Nov 2026.
4. TRL: Transformer Reinforcement Learning. Hugging Face. Accessed 9 Nov 2026.
5. meta-models/Muse-Glimmer-30B. Meta (Hugging Face). Accessed 9 Nov 2026.
6. Llama 4 Community License Agreement. Meta (Hugging Face), 5 Apr 2025. Accessed 9 Nov 2026.

Further reading

• Buyer's guide · Updated quarterlyThe best open-source AI in 2027, for companies that want to own itAn analysis of the open-weight model families a company can run today and the layer that turns one into its own, compared on seven published criteria, with sourced licences, costs and the established path from rented AI to owned AI.
• Licence guideOpen source vs open weights: what each licence lets a company doAlmost every model sold as open-source AI publishes its weights but not its training data. The difference rarely matters for internal use. It matters a great deal for companies that resell a model or build a product on it.
• Practical guideHow to train your own AI model on company dataMost of the work in training a company model happens before any training starts. A step-by-step account of the data work, the choice of base model, the training methods and the rules that apply in Europe.

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski