Meta releases 816B-parameter Muse Spark Open model, runs locally on paired Mac Studios

Meta Platforms has released Muse Spark Open, an 816-billion-parameter open-weight AI model with 42 billion active parameters and a one-million-token context window, designed for coding and agentic tasks such as extended tasks, multi-step reasoning, and software engineering. It is the first model in its class capable of running on a desktop, specifically on two Apple M5 Ultra Mac Studios using Apple's MLX framework. A compressed 4-bit version occupies roughly 493GB, too large for a single 512GB Mac Studio. Apple benchmarked the 4-bit model at 24.99 tokens per second on two M5 Ultra Mac Studios using tensor parallelism, rising to 36.73 tokens per second across four machines, using RDMA over Thunderbolt 5 cables for low-latency data transfer.

Meta CEO Mark Zuckerberg thanked Apple's John Ternus for the collaboration. Alexandr Wang, Meta's chief of AI, called Muse Spark Open "the most capable US open-weight model to date" and joked to Axios about the Apple-Meta partnership. The release ships with custom Helion GPU kernels in SGLang, delivering 1.03 to 3.95 times kernel-level speedups on Nvidia and AMD GPUs, plus a 3% end-to-end speedup on AMD MI350. Model weights are available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models, and accessible via API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, under the Muse Model License with limited commercial terms.

Muse Spark Open follows the earlier 2026 launch of Muse Spark, the closed proprietary flagship powering the Meta AI assistant. The open-weight release is positioned against Chinese competitors including Alibaba's Qwen and Moonshot AI's Kimi, which have dominated the open-weight model market.

(Source：Valdrans)