On Oct 15, 2026 local time, Meta officially released the open-weight model Muse Spark Open, proclaiming it as the most capable U.S. open-weight model to date. Launched shortly after the closed flagship Muse Wonder, the model uses a native multimodal architecture with 816 billion total parameters (42 billion activation parameters) and a 1 million-token context window. It is tuned for code construction and agent collaboration scenarios, showing strong performance on long-range tasks, complex software engineering, and multi-step reasoning. The model weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, and the model is supported by major inference platforms.

Details

Muse Spark Open is built on the earlier Muse Spark series (sharing the same base model as Muse Spark 1.3) and underwent additional safety fine-tuning to improve robustness when running without system-level guardrails. Officially, Meta positions it as a "desktop-class" frontier open model for developers and agent workflows.

Core specifications：

Architecture: Sparse Mixture of Experts (MoE), 816B total parameters, ~42B activated per token.
Context: Native support for a 1 million-token window, capable of containing large codebases, complex dependencies, and thousands of files without splitting tasks.
Multimodal: Native input support for text, images, audio, and video, with text output; includes a perceptual encoder able to directly interpret screen recordings, video clips, or documents.
Optimization focus: Long-range agent tasks, complex software engineering, multi-step reasoning, and tool use.

Performance and benchmarks: Meta's published comparisons show Muse Spark Open leading many U.S. and some open-source models on agent benchmarks such as JobBench and MCP-Atlas—for example, a JobBench score of 61.8, notably higher than some competitors. It excels in tool usage and multi-step collaboration. However, on certain pure coding benchmarks (e.g., Terminal-Bench 2.1) it still lags behind some Chinese open-source models (such as Kimi K3 and GLM 5.3). Overall, it is positioned as the "strongest representative of the U.S. open-source camp," rather than as a universal winner across all global open models.

Deployment and ecosystem：

Supports major inference engines including SGLang, vLLM, TensorRT-LLM, MLX, and llama.cpp, with optimizations for NVIDIA, AMD, Intel, and Apple silicon.
In particular, Meta collaborated with Apple to optimize via the latest MLX framework so the model can run locally on two Mac Studio machines equipped with M5 Ultra, offering a "desktop supercomputer" experience. Official demos include end-to-end driving of DaVinci Resolve for video editing and submitting complete iOS features and fixes to the open-source app Ice Cubes.
Weights are available in BF16 and quantized variants; local runs demand high-end hardware (NVFP4 quantization requires roughly 700GB+ VRAM, BF16 needs more).
Available via API through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
Supports further fine-tuning with PyTorch TorchTitan (including multi-turn SFT, LoRA, and reinforcement learning).

License: Distributed under the Muse Model License, permitting free use, modification, fine-tuning, distillation, and distribution, with limited commercial terms and no additional fees.

Meta AI chief Alexandr Wang emphasized that this move continues the company’s commitment to open models, while noting that "for safety and capability reasons, we generally will not open-