Meta Open-Sources Muse Spark: Open Weights but Not Open Source, with Three Commercial Strings Attached

Meta released the model weights for Muse Spark Open on October 15 — a natively multimodal mixture-of-experts model with 816B total parameters, 42B activated per token, a 1 million token context window, and a focus on coding and agentic tasks. Meta officially calls it "the strongest American open-weight model to date." That claim comes from Meta's own blog, backed by Meta's own benchmark table.

Release Contents and Technical Specs

The weights are live on Hugging Face, available in both BF16 and NVFP4 checkpoints, and also landed on Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted APIs became available the same day through platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Runtime support covers SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. Meta even demonstrated it running on two Apple M5 Ultra Mac Studios, each with 512 GB of memory.

According to the Hugging Face model card, the architecture uses 256 experts per MoE layer with 8 routed per token, supports text, image, audio, and video input, and outputs text.

Why It Matters

For most of 2026, the lead in open-weight models has belonged to Chinese labs like Qwen, Kimi, and GLM. Muse Spark Open is a US lab's response: a large, agent-oriented model you can actually download and self-host.

Mixture-of-Experts (MoE) architecture: Traditional dense models activate all parameters for every token, whereas MoE places multiple parallel "expert" subnetworks in each layer, with a router activating only a few based on token content. Muse Spark Open has 256 experts per layer and routes 8 per token, so the computation path uses only a fraction of the total parameter count. This design lets the model hold enormous knowledge capacity (816B total parameters) while keeping per-token compute roughly equivalent to a 42B model. The trade-off is that all expert weights still need to be stored and loaded, which is the fundamental reason memory requirements far exceed the activated scale.

NVFP4 vs. BF16 precision: BF16 is a 16-bit floating-point format that preserves a relatively high numerical dynamic range and is a common default precision for training and inference; NVFP4 is a 4-bit floating-point microscaling format proposed by NVIDIA that uses shared scaling factors to maintain precision as much as possible at very low bit widths. Lower bit width means less memory for weights and higher inference throughput, but may come with slight quality loss. This explains why the same model needs roughly 700 GB of memory under NVFP4 versus about 2 TB under BF16.

Three Things Most Coverage Overlooked

Activated parameters are not total parameters

The 42B activated parameters determine per-token compute, but all 816B weights still need to reside in memory. The model card specifies roughly 700 GB of memory at NVFP4 precision and about 2 TB at BF16. Per-token inference cost may not be high, but hosting costs are very high.

The license is not an OSI open source license

The Muse Model License starts out reading like MIT ("free... unrestricted") but then attaches conditions:

Section 2: Any product using the model or "derivatives" must prominently display "Built with Muse Spark Open" in the interface, documentation, website, or settings. "Derivatives" here includes any model that "uses this software or its outputs/results for training, modification (including fine-tuning), distillation, or other improvements" — fine-tuning a small model on Muse outputs also counts as a derivative.

Section 3: If you sell model access via API ("Model as a Service") and the group's total revenue exceeds $20 million in any 12-month period, you must first sign a separate agreement with Meta. Internal use, research, and powering features for your own products are not subject to this restriction.

The accompanying usage policy also prohibits using the model to operate critical infrastructure, transportation technology, or heavy machinery, and states the model is not intended for users under 18.

The difference between OSI open source licenses and "open weights": Open source licenses certified by the OSI (Open Source Initiative), such as MIT and Apache 2.0, must allow anyone to freely use, modify, and distribute the software, and must not attach discriminatory restrictions on fields of use or business models. "Open-weight," by contrast, only means the model parameters are downloadable; the license can attach conditions like attribution, revenue thresholds, and usage restrictions. The Muse Model License falls into the latter category: it grants the ability to download and self-host, but imposes constraints on attribution, API resale, and acceptable use, so it should not be treated as open source software.

Derivatives and distillation: Distillation typically refers to using the outputs of a strong model (teacher) as training signals to train a smaller model (student). Under the Muse Model License's definition, any model trained, fine-tuned, or improved using this model's outputs or results counts as a "derivative" and is likewise subject to the prominent "Built with Muse Spark Open" attribution obligation. This clause has a particularly direct impact on small teams doing data distillation or instruction fine-tuning.

Benchmarks are self-reported

Meta says a more complete evaluation report is coming soon. Currently, independent leaderboards show Muse Spark Open trailing the strongest Chinese open models on general intelligence scores. Multiple sources (Meta's announcement, the Hugging Face model card, and the license text) corroborate each other on technical specs, but only Meta's unilateral data backs the capability ranking.

Where Facts End and Predictions Begin

The above content comes from Meta's announcement, the Hugging Face model card, and the license text. The following are personal judgments rather than facts: most teams will use it through hosted APIs rather than local hardware; for small teams running distillation pipelines, the derivatives clause will matter more than the $20 million clause.

Practical Recommendations

1. Read LICENSE.md before building anything on top of it, especially if you plan to distill or fine-tune on its outputs.
2. Add "Built with Muse Spark Open" in your product's interface, documentation, or settings.
3. Before trusting any leaderboard, run 20 to 50 of your own real tasks through a hosted endpoint.

Reference sources:

• Meta AI Research, Introducing Muse Spark Open
• Muse Model License (Hugging Face)
• Hugging Face launch post

(This article was AI-assisted and human-reviewed)

One-line Summary

Meta releases the 816B-parameter open-weight model Muse Spark Open, but the license and self-reported benchmarks hide constraints.

Paragraph Summary

Meta released the Muse Spark Open model weights on October 15, featuring 816B total parameters, 42B activated per token, a 1 million token context, and a focus on coding and agentic tasks, which Meta calls "the strongest American open-weight model to date." The weights are live on Hugging Face and several cloud platforms, with multiple hosted APIs available simultaneously. The article flags three things most coverage overlooked: activated parameters are not total parameters, and all weights still need to reside in memory (about 700 GB at NVFP4, about 2 TB at BF16); the license is not an OSI open source license — using the model or derivatives requires prominent "Built with Muse Spark Open" attribution, and API resale with group annual revenue over $20 million requires a separate agreement; and benchmarks are self-reported by Meta, while independent leaderboards show it trailing the strongest Chinese open models on general intelligence. The article advises developers to read the license first, add attribution, and test with real tasks before trusting leaderboards.