Meta releases Muse Spark Open and teams up with Apple

Contents

Meta has released a new AI model with open weights. The company calls Muse Spark Open the most capable open-weight model from the U.S. to date. It comes just over two months after Muse Glimmer and, with 816 billion parameters, is considerably larger and — according to independent tests — significantly more capable, but it also requires substantially more powerful hardware.

According to its model card, Muse Spark Open uses a mixture-of-experts architecture with 256 experts, eight of which are activated for each token. That means about 42 billion parameters are active per token. The context window holds 1 million tokens. The model processes text, images, video and audio, but outputs only text. In its announcement, Meta pitches it for coding, multistep reasoning and AI agents that use tools, correct errors and handle longer tasks.

Meta and Apple team up

Meta worked with Apple, of all companies, to run Muse Spark Open locally. The two companies have been cooperating for several months, Axios reports. Mark Zuckerberg publicly thanked Apple and CEO John Ternus on X. The partnership is notable because the two companies spent years fighting over privacy and business interests.

According to Apple’s research report, the model variant, reduced to about 530 GB, can run across two connected Mac Studios, each with an M5 Ultra and 512 GB of unified memory. Apple’s MLX framework distributes the computing across both devices over a Thunderbolt 5 connection. At a context length of 8,192 tokens, Apple measured about 25 output tokens per second. However, the required Mac Studios with 512 GB of memory won’t be available until late October and are expected to cost tens of thousands of euros combined.

Lags Chinese competitors on agent tasks

According to tests by Artificial Analysis (AA), Muse Spark Open scores 39 points on the Intelligence Index, making it the most capable open-weight model from the U.S. at present. Nvidia’s Nemotron 3.5 Super VL, released almost simultaneously, scores 26 points, but is considerably smaller, with 120 billion parameters. Leading Chinese models rank higher: MiMo-V2.6-Pro scores 46 points, while the similarly sized GLM-5.3 scores 45.

Muse Spark Open has weaknesses on agent tasks: Among the nine most capable open-weight models, it ranks last overall across four agent benchmarks, but places fourth in each of the other six tests. According to AA, the entire six-point gap behind GLM-5.3 is due to the agent tests.

One strength is how it handles gaps in its knowledge. Muse Spark Open scores 25 points on the AA Omniscience benchmark, the highest score among open-weight models. Its hallucination rate on questions it fails to answer correctly is 24%, the lowest among open-weight models with at least 30 index points.

Low token usage, high API prices

Muse Spark Open generates an average of 40,000 output tokens per task on the Intelligence Index, fewer than any of the other eight leading open-weight models, Artificial Analysis writes on X.

At the API providers available at launch — Together AI, Fireworks and Baseten — 1 million input tokens cost $1.25 and 1 million output tokens cost $4.25. Despite the comparatively high token prices, AA calculates an average cost of $0.84 per benchmark task thanks to the model’s low usage. GLM-5.3 costs $2.01 per task, while DeepSeek V4.1 Flash costs just $0.27.

Anyone who wants to run Muse Spark Open themselves can download the model weights from Hugging Face. The weights are available in a 16-bit version of about 1.6 TB and a version partially quantized to 4 bits that is about 530 GB.

The Muse Model License generally allows commercial use, but requires users to include the notice “Built with Muse Spark Open” and comply with an acceptable-use policy. Companies that offer Muse Spark Open as a service through APIs need a separate agreement with Meta if their gross revenue exceeds $20 million within 12 months. Revenue from affiliated companies also counts.

(dahe)