Meta releases Muse Spark Open, an 816-billion-parameter open-weight model that runs on Mac Studios

Release and specifications

Meta published the weights for Muse Spark Open on 15 October, calling it the "most capable U.S. open-weight model to date". The model has 816 billion total parameters, 42 billion of them active, and a context window of 1 million tokens. Meta built it mainly for coding and agentic work, and its blog post describes the model as natively multimodal. The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta says the model is free to download under a license that allows limited commercial use.

Running it on Mac hardware

Meta worked with Apple to run the model on a desktop setup of two 512GB M5 Ultra Mac Studios using Apple's MLX framework. Each machine's integrated GPU can draw up to 512GB of unified memory, so the pair together offers a terabyte. Apple has not yet shipped the 512GB Mac Studio. Engadget reports that Apple gave a "late October" window for it, while MacRumors says it is expected later this month. Engadget also says the model can run on four 256GB M5 Ultra Mac Studios, which it estimates would cost roughly $38,000. MacRumors notes that the 256GB Mac Studio is already priced above $10,000.

Benchmarks and rivals

Meta's model card gives Muse Spark Open an Artificial Analysis Intelligence Index score of 39. MiMo-V2.6-Pro, which an independent assessment scored at 46, sits higher on that index. Meta's own table places Spark Open ahead of its listed rivals on the JobBench and MCP-Atlas tool-use tests. It trails Kimi K3 and GLM 5.3 in terminal coding, long software-engineering work and GDPVal-AA, a professional-work test. Engadget reports that Meta's benchmarks put the model ahead of GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. VentureBeat says a separate Artificial Analysis assessment of Muse Spark Open could not be verified at publication. The announcement omits hosted token rates, so it does not establish a cost advantage.

A closed frontier and an open second release

Meta said on Monday, 12 October, that Muse Wonder had reached the frontier, but it published no benchmarks and the model stays closed. Alexandr Wang, Meta's chief AI officer, said the company will not open-source its most powerful models.

Because of both safety and capabilities, we will generally not open-source our most powerful models.

Wang said Meta intends to keep releasing good open models. Muse Spark Open is the second open release from the Muse line, after Glimmer, a 30-billion-parameter model distilled from Muse Spark and published under Apache 2.0 in August. In September the company had not decided whether to publish the weights for Muse Spark 1.3.

Regulation and the open-model race

Article 53 of the EU AI Act exempts genuinely free and open models from documentation owed to the AI Office, but the exemption ends at systemic risk. The European Commission can fine companies up to 3% of global turnover. Meta declined to sign the code of practice written to operationalise those duties. Zuckerberg told the Sources podcast about five weeks earlier that rivals training more advanced models and not releasing them was quite dangerous, though he named no company. A Hugging Face assessment from March 2026 found Chinese models accounted for 41% of downloads over the preceding year. In September, OpenRouter reported that open-weight systems made up 60% of token traffic from American-origin requests on its platform.

Apple partnership

Apple and Meta have often been rivals, so their joint work on Muse Spark Open stands out. Wang told Axios that the companies had worked together for several months, with Daniel Gross, a former Apple executive now at Meta, playing a key role. Mark Zuckerberg thanked Apple CEO John Ternus for the collaboration, according to MacRumors.

We're excited for mommy and daddy to be playing nice again.