THE SCOREBOARD FOR WHAT'S NEXTAI + QUANTUM, SCORED. COMPARED. EXPLAINED.
AIQUANTUMSCORE.
← AI & Quantum News

AI / NEWS BRIEF

Meta releases 816-billion-parameter Muse Spark Open weights

Meta has released Muse Spark Open under its own model license. The multimodal mixture-of-experts model has 816 billion total parameters, but practical self-hosting requires unusually large hardware.

By AIQuantumScore · · Event 2026-10-15

What was released

Meta published the weights and model card for Muse Spark Open on October 15. The mixture-of-experts model has 816 billion total parameters, activates 42 billion parameters, and accepts text, images, video and audio while producing text. Meta specifies a one-million-token inference context and says the model was built for coding, tool use, multimodal reasoning and long-running agent tasks. The weights are distributed under Meta's Muse Model License rather than a standard open-source software license.

The hardware threshold

The release is downloadable, but it is not a typical local model. Meta says the NVFP4 version needs at least 700 GB of GPU memory and the BF16 version needs 2 TB. It lists support for common inference frameworks and says an MLX-optimized configuration can run across two M5 Ultra Mac Studios. Third-party API access is also available. Those requirements sharply limit who can self-host the full model despite the open-weight distribution.

Why it matters

OUR ANALYSIS

The release gives researchers and well-equipped organizations direct access to a large multimodal agent model instead of only a hosted endpoint. That makes independent inspection, quantization and specialized deployment more practical, though access terms and infrastructure costs still matter. Meta publishes benchmark comparisons and calls the model highly capable, but AIQuantumScore has not independently reproduced those results. No REVAL or leaderboard score should be inferred from the model card.

More to follow