Meta releases Muse Spark Open, an 816B open-weight model it says runs on two Mac Studios

Meta released the weights of Muse Spark Open on Thursday 15 October 2026, calling it in a blog post "the most capable US open-weight model to date" and "the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework."

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. Meta competes with Anthropic in AI models, though no Claude model appears in Meta's comparison tables for this release.

What Meta released

Meta describes Muse Spark Open as "a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window", "Built for coding and agentic work".

Meta says it is available through seven API providers and as a download from Hugging Face and three cloud platforms. The Hugging Face repository is gated with manual approval; the form asks for name, date of birth, country, affiliation and job title.

It is a different model from Muse Wonder, which Meta released on 12 October and calls its most capable model, available through Meta's API and Muse Code.

The two Mac Studios, and the conditions

Meta's post says: "Apple’s unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model." Apple's UK specification page lists 512GB as the top memory option for the M5 Ultra Mac Studio.

The model card sets out the general requirement: "The model is available in BF16 and NVFP4 versions. To run on your own system, the model requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16." It adds: "It is also optimized to run on two Mac Studios with M5 Ultra using MLX."

Neither the post nor the card states which precision or quantisation the Mac setup uses, and neither gives a speed figure for it. On our reading, the 2TB BF16 version would not fit in two 512GB machines. Community-made 3-bit and 4-bit MLX conversions were posted to Hugging Face on 15 October by the mlx-community account, which is not Meta. The card also notes that "Quantized inference may show minor quality differences in edge cases compared to full-precision."

Benchmarks: two tables

The blog's scorecard is an image. It has 11 rows comparing Muse Spark Open with seven models Meta describes as open-weight, among them GLM 5.3, Kimi K3, Qwen3.8-2.4T-A95B and Mistral Large 4. On our count, Muse Spark Open has the top score on six rows (JobBench, MCP-Atlas, HLE-Diamond without tools, AA-Omniscience, ZeroBench and SimpleVQA). A rival scores higher on five:

• GDPval-AA v2: GLM 5.3 1653, Muse Spark Open 1401
• SciCode: Kimi K3 59.5, Muse Spark Open 58.5
• Terminal-Bench 2.1: Kimi K3 85.0, Muse Spark Open 76.4
• DeepSWE 1.1: Kimi K3 68, Muse Spark Open 60
• BabyVision: Kimi K3 85.7, Muse Spark Open 80.2

The model card's table is longer: 22 rows, 20 capability benchmarks plus two safety rows. On our count, Muse Spark Open has the top score on 7 of the 20 capability rows and a rival is higher on 13. Rows in the card but not the blog image include the Artificial Analysis Intelligence Index (GLM 5.3 45, Kimi K3 44, Qwen3.8 40, Muse Spark Open 39) and GPQA Diamond (Qwen3.8 and Kimi K3 93.5, Muse Spark Open 89.0).

Meta's evaluation methodology report says that for Beam and Mistral Large 4 it cites "their self-reported numbers and results from Artificial Analysis when available". Depending on the benchmark, other rival figures come from Meta's internal runs, Artificial Analysis, Scale AI, official leaderboards or the developers' own reports.

Safety figures

On FORTRESS, Muse Spark Open has the lowest Average Risk Score (5.8, lower is better) of the six models listed, and the highest Over-Refusal Score (17.4), which the methodology report defines as "the percentage of benign prompts that a model refuses to answer (lower is better)". On StrongREJECT v2, which the report calls an internal benchmark, it scores 0.5 against 62.9 to 86.4 for the others listed; the score is an attack success rate, where lower is better. Meta says the model "met the thresholds our framework requires for release".

Licence

One licence covers both code and weights, and Meta's post links no separate code repository. Hugging Face's cardData lists license: other and license_name: muse-model-license; the text grants rights to "the model, code, model weights, parameters, configuration files, and associated documentation files". The MLX adapter Meta links to, Apple's mlx-lm repository, is MIT-licensed according to GitHub. The main conditions:

• Products using the model or any derivative must "prominently display" the words "Built with Muse Spark Open", in the interface or in accompanying documentation, website or settings.
• A licensee offering AI models to third parties via API or similar ("Model as a Service") whose gross revenue, with affiliates, exceeds $20 million in any consecutive 12 months "must notify Meta and enter into a separate written agreement with Meta". The licence says that clause does not apply to internal use, research, other commercial activity that is not Model as a Service, or features of the licensee's own product.
• The Usage Policy applies. Its five headings bar use to break the law or others' rights (including malware, and unlicensed legal, medical or financial practice); activities risking death or bodily harm, including "Military, warfare, nuclear industries or applications, espionage" and ITAR-controlled material; deceiving others, including "Representing that the use of Muse Spark Open or outputs are human generated"; failing to disclose known dangers to end users; and using tools designed for unlawful content.
• "Muse Spark Open is not intended for individuals under the age of 18."

The licence contains no EU or UK exclusion. It names Meta Platforms Ireland Limited as the licensor for anyone in "the European region", and Meta Platforms, Inc. elsewhere.

Why it matters

On our reading, the desktop claim comes with an unstated condition: Meta says two Mac Studios can hold the whole model, but does not say at what precision, while its card puts the NVFP4 version at 700GB of VRAM or more. Its two tables also differ: the blog's 11 rows show Muse Spark Open ahead on six, the card's 20 capability rows show a rival ahead on 13. Meta says a full evaluation report is "available soon".