Meta Releases Muse Spark Open: 816B Open-Weight Model Co-Optimized with Apple for Local Mac Studio Deployment
On October 15, Meta officially released Muse Spark Open, an 816-billion-parameter, 42-billion-active-parameter MoE multimodal model co-optimized with Apple so that two M5 Ultra Mac Studios can run it locally via the MLX framework. Meta describes it as the most capable U.S. open-weight model available today.
The model supports a one-million-token context window and accepts text, image, video, and audio inputs. Rather than generating media, it focuses on deep cross-modal understanding, making it especially useful for coding, multi-step reasoning, and long-horizon agentic workflows. It has been adapted for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, and optimized for AMD, Apple, Intel, and NVIDIA hardware.
The most distinctive feature of the release is the Apple collaboration. Two M5 Ultra Mac Studios, each with up to 512 GB of unified memory, are linked through the MLX framework to behave like a single desktop supercomputer with one terabyte of shared memory. This lets developers run inference entirely on local hardware, keeping sensitive prompts, code, and enterprise data off the cloud while reducing long-term reliance on rented compute.
Meta's own benchmarks place Muse Spark Open ahead of GLM 5.3, Kimi K3, and Qwen 3.8 on JobBench and MCP-Atlas tool-use tests. The weights are available on Hugging Face, ModelScope, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, while API access is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
The launch signals a broader industry shift: open-weight models are closing the gap with closed frontier systems, and the combination of high-memory local hardware plus open models is becoming a viable private-deployment option for enterprises. The high hardware cost means mass adoption will depend on future quantized releases, but the precedent has been set.
Sources: Mashable SEA, Tencent News, 163.com, Sina Finance, Yahoo Finance; threerouter
Top comments (0)