Meta has released the weights for Muse Spark Open, describing it as the most capable US open-weight model to date. The headline numbers are substantial: 816 billion total parameters with 42 billion active, a 1-million-token context window, and native multimodality.
What makes it different
Open-weight releases with large parameter counts are not new. What is new here is the stated deployment target. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the MLX framework.
The mechanism is Apple’s unified memory architecture. The M5 Ultra’s integrated GPU can address up to 512 GB of memory per Mac Studio, and the chip’s memory bandwidth is rated at 1.2 TB/s. Two machines can therefore hold an entire model that would otherwise need a server rack.
That changes who can run a frontier-class model locally. It does not make it cheap — two maxed-out Mac Studios is a serious purchase — but it removes the need for a data centre relationship.
The specs that matter
- 816B total parameters, 42B active: the mixture-of-experts structure means only a fraction of the model runs per token, which is what makes local inference plausible at all.
- 1M-token context window: enough for large codebases or long document sets in a single pass.
- Natively multimodal: text and images are handled in the same model rather than bolted together.
- Toolchain support: Meta lists Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA silicon. The release also ships custom Helion GPU kernels in SGLang, with 1.03-3.95x kernel-level speedups on NVIDIA and AMD GPUs.
The licence is not fully open
This is the part worth reading carefully. Meta is releasing the weights under the Muse Model License with “limited commercial terms,” which allow many developers to use the model for free. That is a source-available arrangement, not an unrestricted open-source licence. If you plan commercial use, read the terms rather than assuming.
Meta also states the model was evaluated against its Meta Superintelligence Scaling Framework, covering the most severe risk categories, and met the release thresholds. Detailed results are in the model card.
What we did not do
We did not download the weights, run the model, or benchmark it. The parameter counts, context window, memory figures and speedup numbers above are Meta’s own published claims. Treat performance claims from any model vendor as claims to verify against your own workload.
Meta also introduced Muse Wonder and Muse Wonder Full as frontier agent models for its API and Muse Code, and we covered the earlier Muse and Apple collaboration and Google’s Gemini 4 Argon rollout.
Sources
- Meta AI Research, “Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date” (October 15, 2026): https://research.meta.ai/blog/muse-spark-open
- Meta, “Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents” (October 12, 2026): https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- Muse Spark Open model weights on Hugging Face: https://huggingface.co/collections/meta-models/muse-spark-open
