AI news · October 16, 2026
Meta releases Muse Spark Open with 816B parameters and a 1M-token context
Meta released an 816B open-weight model with 42B active parameters and a 1M-token context, so builders can test long multimodal tasks on hosted providers or large local machines.
- total parameters
- 816B
- active parameters
- 42B
- context tokens
- 1M
Meta released Muse Spark Open on October 15 as an open-weight multimodal model with 816 billion total parameters and 42 billion active per token. Meta lists a 1,048,576-token context, text and image input, and weights for Hugging Face, SageMaker, Model Garden and Foundry. The model is also available through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Meta says it was run on two M5 Ultra Mac Studios with 512GB each, using MLX.
The comparison is unusual: the active count is close to other sparse models, but the total size and memory need are much higher than a normal developer workstation. The license also limits commercial use, so an API trial is safer than building a paid product around the weights. The context limit is a claim to measure on real files, not proof that quality remains steady at one million tokens.
What you can do with it
Try it through the hosted providers listed in Meta's developer page, or download the weights when your hardware fits. Use a 100-page document set with images and tool calls; record answer quality, time, memory use and the license terms before choosing it over a smaller model.
Our take
This is important for long-context and multimodal experiments, not an easy local default. The 42B active count helps, but a 1TB-class machine and a limited commercial license narrow the first product use. Hosted inference should win the first paid test.
Source: Meta AI Research — Made With Models writes the brief; the reporting is theirs.