Meta Muse Spark Open is an 816B open-weight AI model for developers building long-horizon coding agents and multimodal workflows. Meta released the weights on October 15, 2026, calling it the most capable US open-weight model to date.
Core Features
- 816 billion total parameters with 42.7 billion active per token through a sparse mixture-of-experts design.
- A 1,048,576-token context window that holds entire codebases, dependency trees, logs, and task history.
- Omnimodal input: text, image, video, and audio tokens share one context, with text output.
- Co-trained with Muse Code and tuned for agentic loops that plan, act, observe, and revise.
- Ships in BF16 (about 1.65 TB) and NVFP4 (about 537 GB) checkpoints; runs on a single 4-GPU Blackwell node or two M5 Ultra Mac Studios via MLX.
Use Cases
- Teams self-hosting a frontier-class model to avoid per-token API fees on heavy agent traffic.
- Coding agents that debug from screenshots, reproduce bugs from screen recordings, or take voice-driven instructions.
- Enterprises keeping sensitive prompts and source code inside their own infrastructure.
Pricing
Released under the Muse Model License. Weights are free to download from Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry. Hosted inference is pay-per-token through Baseten, Fireworks AI, Groq, Modal, Together AI, and OpenRouter. NVIDIA reports up to 50,000 tokens per second per GPU on its GB300 NVL72 for agentic workloads.
Pros and Cons
Pros: largest US open-weight release to date, million-token context, native multimodal input.
Cons: heavy serving footprint (537 GB NVFP4 minimum) and a license that requires attribution plus a revenue threshold for Model-as-a-Service.
Our Take
Best for AI engineers who want a Claude-Opus-class backbone without closed-model API fees. The trade-off is the hardware bill: this is a datacenter model, not a laptop download. Browse more in our AI models and engines category.
