Meta Releases Muse Spark Open Weights, Built With Apple to Run on Mac Studio Desktops

Meta on October 15, 2026 released the weights of Muse Spark Open, an 816 billion parameter model for coding and agent work that it calls the most capable US open weight model to date. Meta says it worked with Apple so the model runs on two linked Mac Studio desktops with M5 Ultra chips, using Apple's MLX framework.

The release comes three days after Meta launched Muse Wonder and Muse Wonder Full, frontier models it keeps closed, and sets Meta against the large open weight models from Chinese labs such as Qwen, GLM and Kimi. Muse Spark Open shares its base model with the closed Muse Spark 1.3, according to Meta's PyTorch team. In September, Meta had not decided whether to publish that model's weights; its only open Muse model was Muse Glimmer, a 30 billion parameter model released in August under Apache 2.0 (The Next Web).

What exactly did Meta release?

Meta released Muse Spark Open on October 15, 2026 as a mixture of experts model with 816 billion total parameters, 42 billion of them active for each token, and a 1 million token context window, according to its model card. A mixture of experts model splits its layers into many smaller expert networks and runs only a few per token, here 8 of 256.

It takes text, image, video and audio input and returns text. Meta publishes full precision BF16 weights and a smaller NVFP4 version, a 4 bit format built for Nvidia's Blackwell chips. On GPU servers, the model card asks for at least 700 GB of graphics memory for NVFP4 and 2 TB for BF16.

The weights come under a new Muse Model License with what Meta calls limited commercial terms. The license text grants free use, modification, distillation and resale, but every product built on the model or its derivatives must display "Built with Muse Spark Open" in its interface or documentation. A company that sells access to models through an API, and whose group revenue tops $20 million in any 12 month period, must notify Meta and sign a separate agreement first. Internal use, research and features inside a company's own product are exempt from that clause.

How does Muse Spark Open compare with Qwen, GLM and Kimi?

In Meta's own benchmark table, Muse Spark Open scores higher than Qwen3.8-2.4T-A95B on 8 of the 13 rows where all four models have a score, but higher than GLM 5.3 on only 4 and Kimi K3 on 5. It leads all three on JobBench, MCP-Atlas, HLE-Diamond and AA-Omniscience, and trails all three on 5 rows, including the Artificial Analysis Intelligence Index and Terminal-Bench 2.1.

On the three coding rows, GLM 5.3 and Kimi K3 both score higher. On the seven rows reported only for Kimi K3, Muse Spark Open is ahead on 3 and behind on 4. Meta's methodology report says it used official leaderboards where they exist, otherwise its own best effort runs or vendors' published numbers; Reflection AI's Beam and Mistral Large 4 appear with their makers' own figures.

On the StrongREJECT v2 jailbreak test, where lower is better, Meta reports 0.5 for Muse Spark Open against 62.9 to 66.6 for the three rivals, after what the model card calls additional safety post-training.

What does it take to run Muse Spark Open on a Mac?

Running the NVFP4 version of Muse Spark Open locally takes about 1 TB of unified memory: two M5 Ultra Mac Studios with 512 GB each or four with 256 GB each, linked over Thunderbolt 5, according to Meta's developer blog. Apple says the NVFP4 weights alone take about 493 GB, too much for one machine once working memory is added.

In a technical post, Apple measured 15.74 tokens per second on two Mac Studios that each run half of the layers, 24.99 when two machines split every layer, and 36.73 on four machines, all with an 8,192 token prompt.

The 512 GB Mac Studio behind Meta's two machine setup is not on sale yet; Apple plans it for later in October (CNBC, MacRumors). A Mac Studio with M5 Ultra and 256 GB costs at least $10,799, according to CNBC, which works out to at least about $43,000 for the four machine setup that can be ordered now. Engadget put that setup at roughly $38,000.

The companies worked on the project for months, CNBC reports, after years of conflict that deepened when Apple's 2021 privacy changes hit Meta's ad business.

How does Muse Spark Open compare with Meta's closed models?

Artificial Analysis scores Muse Spark Open at 39 on its Intelligence Index, below the 48 it gave the closed Muse Spark 1.3 and the 52 and 56 it gave Muse Wonder and Muse Wonder Full. The firm tested Muse Spark Open at its xhigh reasoning setting and the closed models at max.

Wang also said Meta will generally not open source its most powerful models, citing safety and capabilities (The Next Web). Hosted, Muse Spark Open costs $1.25 per million input tokens and $4.25 per million output tokens on OpenRouter and Together AI. That matches the price Artificial Analysis lists for the closed Muse Spark 1.3, against $4 and $20 for Muse Wonder.

What we don't know yet

• The price and exact ship date of the 512 GB Mac Studio.
• When Meta's full evaluation report appears; the research blog says it will be available soon.
• What the separate agreement for API providers above the $20 million revenue threshold requires; the license does not set its terms.
• Whether Apple and Meta will work together beyond this model; both declined to say, according to CNBC.