Meta and Apple bring 816B Muse Spark Open to Mac Studio desktops

Meta has released Muse Spark Open, an 816-billion-parameter open-weight AI model for coding and long-running agents. Announced on October 15, the release also marks an unusual collaboration with Apple: the model has been optimized to run locally across two Mac Studio desktops equipped with M5 Ultra chips.

An 816B model that activates only part of its network

Muse Spark Open uses a mixture-of-experts architecture with 816 billion total parameters and roughly 42 billion active parameters per token. The model supports a context window of up to one million tokens. It accepts text, images, video and audio as input and generates text, allowing it to work with documents, software interfaces and multimodal material in a single agent workflow.

Meta is positioning the model for software engineering, multi-step reasoning and tool use. The company describes it as its most capable U.S.-developed open-weight model, a claim about its position among American open-weight releases rather than an assertion that it leads every global benchmark.

Why the Apple collaboration is unusual

Apple worked on MLX-based inference for Muse Spark Open on Mac Studio hardware. Two M5 Ultra machines can provide up to one terabyte of combined unified memory, allowing a quantized version of the large model to operate without sending prompts to an external inference provider. The machines work together rather than running the full model on an ordinary single desktop.

Apple reports that a cluster of four M5 Ultra Mac Studios generated more than 36 tokens per second in a benchmark using an 8,192-token context. The company also notes the memory cost: a four-bit model occupies about 493 GB before intermediate computations, so even the quantized version cannot fit on one Mac Studio.

Developers can download the weights now

Meta has published the model weights on Hugging Face in BF16 and NVFP4 formats. It also supports inference stacks including MLX, vLLM, SGLang, Transformers, llama.cpp and TensorRT-LLM. Hosted access is available through providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

This is an open-weight release, not an unrestricted open-source license. Under the Muse Model License, developers can download and modify the model, but certain conditions apply. In particular, a business with more than $20 million in aggregate revenue over a consecutive 12-month period must arrange a separate agreement with Meta if it offers the model itself to third parties as a model-as-a-service product. That condition does not apply in the same way to ordinary internal use or a product merely powered by the model.

Performance and remaining limits

Meta emphasizes agentic and multimodal benchmarks, but its own model card shows a mixed competitive picture. On the Artificial Analysis Intelligence Index listed in the card, Muse Spark Open scores 39, compared with 44 for Kimi K3 and 45 for GLM 5.3. Results depend on test configuration and should not be confused with performance guarantees for production workloads.

The model also needs substantial computing resources for self-hosting. Running it locally on two premium desktops is technically different from running an assistant on a typical laptop. Developers who do not have access to that hardware can use hosted inference services instead.

Why it matters

The collaboration connects two trends: more capable downloadable models and growing interest in keeping sensitive AI workflows under the user’s control. It also illustrates an unexpected partnership between Meta and Apple, companies better known for competing across consumer platforms and privacy policy.

Sources: Meta AI Research , Apple Machine Learning Research and the official Muse Spark Open model card and license .

Source: Meta AI Research