Muse Spark Open: Meta's open-weight model that runs on Mac Studio
On October 15, 2026, Meta released Muse Spark Open, described in the company's official blog as "the most capable US open-weight model to date". It is a Mixture-of-Experts model with 816 billion total parameters, 42 billion of them active, 256 experts and top-8 routing. Context goes up to one million tokens, and the Hugging Face model card lists it as natively multimodal. Meta pitches it for coding and long-running agentic tasks, such as software engineering and multi-step reasoning.
The Muse Model License is proprietary, not open source: access to the weights is gated on Hugging Face and comes with a Usage Policy (Exhibit A). Anyone who exceeds $20 million in gross revenue over 12 consecutive months by offering the model via API to third parties must notify Meta and sign a separate agreement; internal use, research and integration into your own products are excluded. The license also states that the model is not intended for people under 18 and that Meta claims no ownership of the outputs.
The weights can be downloaded from platforms such as Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, while APIs are hosted on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. For local use, Meta worked with Apple to optimize the model for the MLX framework, so it can run on two Mac Studios with M5 Ultra chips, each with up to 512 GB of unified memory: according to Meta, it's the first model in its class to do so. The model card calls for at least 700 GB of GPU memory in NVFP4 and 2 TB in BF16; TechCrunch, citing Engadget, instead mentions "at least 1 TB of RAM" for local use.
In the benchmarks published by Meta (xhigh reasoning), it leads some tests, such as MCP-Atlas (89.2) and HLE-Diamond without tools (26.1). In others it trails Chinese open models: on Terminal-Bench 2.1 it scores 76.4 against 85.0 for Kimi K3 and 83.9 for GLM 5.3, and on the AA Intelligence Index 39 against GLM 5.3's 45. Worth stressing: all benchmarks come from Meta and have not been independently verified, as TechCrunch pointed out. On safety, Meta added dedicated post-training, evaluated the model with the Meta Superintelligence Scaling Framework and plans to publish a full report soon. The model card recommends using the model within systems with adequate guardrails and human confirmation for irreversible actions.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read Meta's official blog (date, architecture, distribution, Apple collaboration on MLX, safety) and the Hugging Face model card (license, $20 million threshold, hardware requirements, benchmarks, safety notes). TechCrunch's October 15, 2026 report independently confirms the parameter count, the 1M context, running on a Mac and the fact that the benchmarks are internal only. I checked that this doesn't overlap with our earlier pieces on Muse Wonder and on Muse Spark in math: this one is about the weights release.
- Incertezze
- All benchmarks are Meta's, with no independent verification (TechCrunch notes this too). I read the model card table in flattened form: beyond the numbers quoted, column alignment should be double-checked. The full safety report is not out yet. Sources disagree on memory for local use: TechCrunch, citing Engadget, says "at least 1 TB of RAM", while Meta gives 2×512 GB on Mac Studio and 700 GB of GPU memory in NVFP4, so each figure is attributed to its source. It's not known whether Meta and Apple have a commercial deal beyond the MLX optimization. "Open" in the name doesn't mean open source: the license has thresholds and conditions.
- Perché pubblicarla
- It's the largest open-weight model released by a US big tech company in this period. Three things matter for readers: a license that isn't open source, what "runs on a Mac" really means in hardware terms, and an honest comparison with Chinese open models, which stay ahead on several benchmarks.