The report says at least two Mac Studios equipped with

Article

Rare collaboration between Meta and Apple: 816000000000 The __BIBIBI_NUM_0__-parameter Muse Spark is open-sourced and can run locally with two Mac Studios

Meta open-sourced the Muse Spark Open model weights and collaborated with Apple to enable this 816000000000-parameter large model to run locally on Mac Studio. Each inference activates only 42000000000 parameters and supports 1000000 Token context, primarily targeting programming and complex Agent tasks. The weights were listed on Hugging Face on October 15; developers can apply to obtain them after signing in and accepting the access conditions.

Meta and Apple have clashed repeatedly in the past over privacy policies, but Zuckerberg has now publicly thanked Apple's CEO John Ternus, who also responded personally. The two sides use Apple's MLX framework to distribute the model across multiple Mac Studios. Even compressed to 4-bit precision, the model still requires approximately 493GB of memory; with inference overhead, it cannot fit in a single Mac Studio.

Actual deployment requires at least two Mac Studios equipped with M5 Ultra chips, each with 512GB of unified memory, with hardware costs exceeding 40000 USD. Apple's testing shows that four M5 Ultra Mac Studios running together can generate approximately 8192 Token tokens per second with 37 context.

Meta uses a custom Muse Model License that allows developers to modify, commercially use, and distribute the model, but requires them to state in the product or related documentation that Muse Spark Open is used. If an enterprise and its affiliates have total revenue exceeding Token for 12 consecutive months, they must separately sign an agreement with Meta before offering the model externally as an 20000000 USD service.

Original link API https://m.theblockbeats.info/flash/372118

Key points

The report says the Muse Model License allows developers to modify, commercially use, and distribute the model, but requires them to state in the product or related documentation that Muse Spark Open is used.
The report says that if an enterprise and its affiliates have total revenue exceeding 816000000000 for __BIBIBI_NUM_23__ consecutive months, they must separately sign an agreement with Meta before offering the model externally as an 42000000000 service. 1000000 Token The report says Meta and Apple collaborated to open-source Muse Spark Open and enable it to run locally across multiple Mac Studios.
Local deployment Agent Apple chips
If the report's contents are accurate, they demonstrate how an extremely large model can run on a cluster of local devices, and also show that hardware costs and licensing conditions affect actual use.October 15 Unified memory
This is a report about local deployment of a large model. Although the model has a very large parameter scale, each inference activates only a portion of it; running it nevertheless requires multiple high-memory Mac Studios to work together.MLX The key points are: the model is very large, requires multiple high-memory devices, and additional licensing conditions must be observed when commercially using it or providing
services. 4 All of the above collaboration details, model specifications, hardware requirements, operating speed, and license terms come from the supplied report and have not been independently verified against other sources. 493GB The evidence does not provide original announcements or the full license text from Meta, Apple, or Hugging Face.
API M5 Agent 512GB Token 40000 USD。
unverified M5 PROJECT_DISCOVERY 8192 Token single_report_source 37 official_sources_missing Token。
specifications_unverified
12 20000000 USD API

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a report about local deployment of a large model. Although the model has a very large parameter scale, each inference activates only a portion of it; running it nevertheless requires multiple high-memory Mac Studios to work together.

Why it matters to readers

If the report's contents are accurate, they demonstrate how an extremely large model can run on a cluster of local devices, and also show that hardware costs and licensing conditions affect actual use.

API

Risks and unknowns

All of the above collaboration details, model specifications, hardware requirements, operating speed, and license terms come from the supplied report and have not been independently verified against other sources.
The evidence does not provide original announcements or the full license text from Meta, Apple, or Hugging Face.
The evidence does not specify Zuckerberg’s full name or his specific role.

The report says at least two Mac Studios equipped with

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center