Meta launches Muse Spark Open as an 816-billion-parameter open-weight model built for coding, multimodal reasoning and long-running agents. Released October 15, it pairs a one-million-token context window with Apple’s MLX framework, allowing the quantized model to run locally across multiple M5 Ultra Mac Studios.

The release has four defining elements:

  • 816 billion total parameters with 42 billion active
  • Text, image, video and audio inputs
  • Local inference across two or four Mac Studios
  • Downloadable weights under Meta’s custom license

 

Background Reading

 

Meta Launches Muse Spark Open With 816 Billion Parameters

Muse Spark Open uses a mixture-of-experts architecture. Although the model contains 816 billion parameters, it activates 42 billion for each token, reducing the computation required at inference compared with activating the full network. Meta’s model card says its 256 experts route each token through eight experts.

 

The model accepts text, images, video and audio, while producing text output. Meta positions it for software engineering, sequential tool use, multimodal analysis and other agentic work that must preserve a plan over many steps. Its perception encoder processes non-text inputs before the language model reasons over them.

 

A one-million-token context window gives developers room for large codebases, lengthy documents and extended interaction histories. That capacity does not guarantee reliable long-horizon execution: tool scaffolding, retrieval quality, error recovery and application permissions remain decisive in a deployed agent.

 

Meta released both BF16 and four-bit NVFP4 weights. The company says the model is multilingual, with training data spanning more than 100 languages, but warns that it has not evaluated every language represented in that data and that performance may weaken outside its strongest set.

 

Apple MLX Splits Muse Spark Open Across Mac Studios

A Mac Studio with M5 Ultra provides 512 GB of unified memory and 1.2 terabytes per second of memory bandwidth. Apple says Muse Spark Open needs more than 1.6 TB for parameters in native BF16 precision, while four-bit quantization reduces the model footprint to about 493 GB.

 

The quantized weights nearly fill one machine before intermediate inference data is counted, so the full model does not fit on a single Mac Studio. MLX distributes it across multiple systems connected by Thunderbolt 5, using remote direct memory access to move data between the machines with relatively low latency.

 

Two Macs can split the model by depth through pipeline parallelism. Four can divide each layer through tensor parallelism, increasing throughput at the cost of more communication. In Apple’s 8,192-token test, two-machine pipeline inference reached 15.74 tokens per second, while four-machine tensor inference reached 36.73 tokens per second.

 

Those measurements are controlled benchmarks, not guarantees for every workload. Context size, quantization, tool calls and concurrent requests will affect real performance. Even so, the demonstration shows that a frontier-scale model can operate in an office without sending prompts to a hosted inference service.

 

Open Weights Expand Access but Hardware Remains Demanding

Developers can download Muse Spark Open through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is also available from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

 

Meta lists support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth and Apple MLX, with optimizations covering AMD, Apple, Intel and Nvidia hardware. The company says custom GPU kernels improve specific operations, although end-to-end gains can be much smaller than kernel-level speedups.

 

The weights use Meta’s Muse Model License rather than an unrestricted open-source license. Most development, modification and commercial use is permitted, but model-as-a-service businesses with more than $20 million in aggregate annual revenue must notify Meta and obtain a separate agreement before offering the model or derivatives through an API.

 

Local deployment can give organizations more control over prompts, source code and private documents because inference stays on their own machines. The entry cost remains high, however: Meta’s model card specifies at least 700 GB of GPU memory for the NVFP4 version outside the Mac configuration and roughly 2 TB for BF16.

 

Muse Spark Open Safety Claims Need Deployment Tests

Meta says the released checkpoint received additional safety post-training and met its framework’s thresholds across content safety, agentic risk, privacy, cybersecurity, chemical and biological misuse, and loss-of-control evaluations. The model card still recommends application-level guardrails and human confirmation before irreversible actions.

 

Open weights let developers modify the checkpoint, which can weaken safeguards or change behavior beyond Meta’s tests. Local execution also removes the monitoring and policy enforcement that a hosted provider may apply. Organizations will need their own red-team testing, access controls, logging and isolation for tools that can alter files or external systems.

 

Meta’s evaluation report combines official leaderboards, internal runs and some competitor self-reported results. It also acknowledges that common prompts and tool setups may not reflect every rival model’s best configuration. Independent reproduction will therefore matter more than rank claims attached to the launch.

 

Muse Spark Open broadens the range of places where a frontier-scale model can run, but it does not turn a typical desktop into an AI server. Its immediate significance is narrower and more practical: Apple and Meta have demonstrated private, local inference for a very large agentic model using commercially available workstation hardware.