AnswersPublished24 sources
What Meta’s Muse Spark Open Can—and Can’t—Do
Meta released Muse Spark Open’s weights on October 15, 2026, calling it the most capable U.S. The multimodal model has 816 billion total parameters, 42 billion active per token and a 1 million token context window.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Meta released Muse Spark Open’s weights on October 15, 2026, calling it the most capable U.S.
- The multimodal model has 816 billion total parameters, 42 billion active per token and a 1 million token context window.
- Muse Spark Open is not the same as Meta’s proprietary Muse Wonder, and reports differ on the open model’s commercial use terms.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Meta announce when it released Muse Spark Open’s weights on October 15, how does its claim to be the most capable U.S. open-weight. Article summary: On October 15, 2026, Meta released Muse Spark Open’s weights, calling it its “most capable US open-weight model to date,” and announced work with Apple to run it locally on connected M5 Ultra Mac Studios using MLX. [9][7. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
Meta released the weights for Muse Spark Open on October 15, 2026, describing it as the most capable U.S. open-weight model to date. The release also highlighted a collaboration with Apple to run the model locally on Mac Studio computers using Apple’s MLX framework. Those are significant claims, but the available comparisons point to a more specific conclusion: Muse Spark Open performs well on some agentic and tool-use tests, while rivals lead on several coding and broader-work evaluations. 10
19
What Muse Spark Open is
Meta describes Muse Spark Open as a natively multimodal model with 816 billion total parameters, 42 billion active per token and a 1-million-token context window. The weights were released for download, and the model is intended for coding and agentic work, including longer tasks and multistep reasoning. 10
9
The “most capable U.S. open-weight model” label is Meta’s claim, not a claim that the model is the best open-weight system worldwide or at every task. Benchmark results vary by evaluation, and a lead on one test does not establish broad superiority. 10
19
How it compares with GLM 5.3 and Kimi K3
In VentureBeat’s account of Meta’s comparison table, Muse Spark Open leads the listed rivals on JobBench and the MCP-Atlas tool-use test. But it trails GLM 5.3 and Kimi K3 on terminal coding, extended software-engineering tasks and the GDPVal-AA professional-work evaluation. The same report notes that competitors lead on other measures as well. 19
The practical takeaway is a mixed benchmark profile: the model appears competitive for selected tool-using and agentic tasks, but the available results do not support a blanket claim that it outperforms GLM 5.3 or Kimi K3. These are reported benchmark comparisons, not guarantees of performance on a particular developer’s workload. 19
Muse Spark Open is not Muse Wonder
Muse Spark Open is the open-weight release; Muse Wonder is described in reporting as a separate, proprietary frontier model introduced earlier that week. One report says Muse Spark Open falls short of Meta’s paid Muse Wonder Full models on several benchmarks. The excerpts available here do not provide enough comparable results to quantify the gap, and they do not establish that Meta itself publicly acknowledged that specific limitation. 21
29
That distinction matters when interpreting the release: downloadable weights offer a way to run and adapt a model, but they should not be treated as evidence that Meta has opened access to its entire frontier model line. 21
22
Running it on Apple Mac Studios
Meta and Apple say Muse Spark Open has been optimized with MLX to run locally on Mac Studio computers. Reporting describes a two-machine setup using two M5 Ultra Mac Studios, each with 512GB of memory, connected via Thunderbolt 5. Meta has also been reported as supporting configurations of two or four machines. 10
2
8
The hardware requirement is substantial. One report puts the compressed NVFP4 model weights at roughly 493GB before the additional memory needed during inference. That is why the reported setup uses multiple high-memory machines rather than a typical desktop computer. 7
2
Local inference means prompts can be processed on the developer’s machines rather than sent to Meta’s model-serving service for that inference. It is not, by itself, a complete privacy guarantee: network-connected tools, telemetry or other software may still transmit data. 4
10
Download access and commercial-use caveats
Reporting says the weights were available through Hugging Face on release day, and Meta’s announcement describes them as released. 10
15
Reports do not describe the commercial terms consistently. CNBC characterizes the license as allowing limited commercial use, while other coverage describes commercial use more broadly and reports a separate agreement requirement for some model-serving businesses above a revenue threshold. Since the excerpts here do not reproduce the full license, don’t assume that every commercial use is permitted; review the license terms for the intended use. 4
12
14
15
What the Apple-Meta partnership signals—and what it doesn’t
The companies reportedly worked together for months, with former Apple executive Daniel Gross, now at Meta, identified as one of the people who helped broker the collaboration. The launch marks a public instance of cooperation between companies that have also competed over devices and disagreed over privacy and computing. 12
13
The technical significance is clearer than any broader business implication: this release demonstrates a local deployment path for one large open-weight model across multiple Mac Studios. It does not, by itself, establish that the setup will cost less than other options, match production-scale accelerator performance or lead to an ongoing Apple-Meta partnership. 8
10
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Meta’s Muse Spark Open Can—and Can’t—Do"?
Meta released Muse Spark Open’s weights on October 15, 2026, calling it the most capable U.S.
What are the key points to validate first?
Meta released Muse Spark Open’s weights on October 15, 2026, calling it the most capable U.S. The multimodal model has 816 billion total parameters, 42 billion active per token and a 1 million token context window.
What should I do next in practice?
Muse Spark Open is not the same as Meta’s proprietary Muse Wonder, and reports differ on the open model’s commercial use terms.