Muse Spark Open

Muse Spark Open (15 October 2026) is Meta's open-weight release of its Muse Spark model — a mixture-of-experts with about 816 billion total and 42 billion active parameters, a context of up to 1M tokens, and text, image, audio and video input with text output. Meta positions it for coding and agentic work.

Best for

• Teams that need to run a frontier-class model on their own infrastructure
• Coding agents and long debugging sessions
• Fine-tuning on private data (Hugging Face shows a LoRA recipe)
• Multimodal input — screenshots, audio and video alongside text

How it compares — and the India angle

• "Open" here means open weights under Meta's own Muse Model License, not an open-source licence — read the commercial terms before shipping it in a product.
• Self-hosting is serious hardware: Hugging Face's guide uses one B200 node for the NVFP4 weights and four 8xH200 nodes for BF16. Most Indian developers will use it through a hosted endpoint instead.
• Independent results are mixed: Vals.ai places it 38th of 49 models on its index. Meta's own claim — the most capable US open-weight model — is a claim about US models, not all open models.
• If data residency is the reason you want open weights, it now competes with DeepSeek V4.1 Flash (MIT licence) and Qwen 4.