Meta ships its first frontier model, Muse Wonder, then open-sources Muse Spark

AIAI #frontier-models #meta #open-weights

Meta ships its first frontier model, Muse Wonder, then open-sources Muse Spark

Meta's own posts and TechCrunch's report show both halves of its strategy in one week: a paid frontier model and a big open-weight one.

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first attempt at a frontier model. By Meta's own testing they score comparably to Fable 5.1 and GPT-6 Astra, and Meta highlights "active memory", recall of what a user has told the model over weeks or months. Wonder costs $4 per million input tokens and $20 per million output tokens through Meta's API; Wonder Full is only for Muse Code subscribers.

On Thursday Meta also released the weights for Muse Spark Open, an 816-billion-parameter (42 billion active) multimodal model with a 1-million-token context window, which it calls the most capable US open-weight model so far. It is offered through hosting providers and on Hugging Face, and Meta says it runs on two Apple M5 Ultra Mac Studios.

Read it at the source

TechCrunchMeta releases its first frontier model, Muse Wonder Independent ↗ Meta AIIntroducing Muse Spark Open, The Most Capable US Open-Weight Model to Date Independent ↗ Hacker NewsDiscussion: Muse Spark Open Independent ↗

Cascadic Analysis 3

UndertowBenchmark scores can overstate real-world reliability

What hidden risk could pull against this, even if the news is good?

The comparisons to Fable 5.1 and GPT-6 Astra, and the lead on MemoryBench, come from Meta's own testing in its model card. Memory over "days, weeks, months" is exactly the kind of skill a benchmark run can't fully show, because it only fails after real use has piled up.

UndertowPersistent memory creates a new poisoning surface

What hidden risk could pull against this, even if the news is good?

Meta's headline feature is "active memory": the model keeps what you tell it and uses it later. Anything that gets into that memory, including a planted instruction or a wrong fact from a web page, can keep shaping answers long after the conversation that put it there is forgotten.

UndertowLoss of meaningful human control as capability and autonomy increase

What hidden risk could pull against this, even if the news is good?

Muse Spark Open's weights are now downloadable and run on two Mac Studios. Once an open-weight model is out, nobody, including Meta, can recall it, patch it, or see how it is being used or modified.

Rabbit Holes 1

Open-source artificial intelligence

After exploring Meta releases its first frontier model, Muse Wonder, the definition to hold it against is Open-source artificial intelligence.

The Open Source Initiative's definition: an AI system free to use, study, modify and share, including the training data, code and model parameters, so someone could build a substantially similar result.

Wikipedia · Swim · 30 min

0

Meta ships its first frontier model, Muse Wonder, then open-sources Muse Spark

Further down the stream