Meta has released Muse Spark Open, a multimodal Mixture-of-Experts model featuring 816 billion parameters and a 1 million token context window. The architecture includes 62 layers, a hidden dimension of 8192, and 256 experts per layer, with 8 experts activated per token. Visual processing is managed by a dedicated 1.8 billion parameter Perception Encoder, while the vocabulary supports over 100 languages with 202,048 tokens.
The model is available in BF16 and NVFP4 formats on Hugging Face, with weights requiring 1.6 TB and 530 GB respectively. Recommended VRAM for deployment is 2 TB for BF16 and 700 GB for NVFP4. Meta has included additional safety post-training to enhance robustness. The Muse Model License allows for fine-tuning, distillation, and redistribution, with exemptions from revenue clauses for internal research and product integration.
Performance benchmarks on Terminal-Bench 2.1 show a score of 76.4, trailing competitors like Kimi K3 and GLM 5.3. The model supports reasoning effort settings, with xhigh as the default. To accelerate decoding, Meta introduced DFlash V2 and DSpark draft models, reporting a 3.7x throughput increase on AMD MI355X hardware using DFlash V2.
Integration support is available via engines including vLLM, SGLang, TensorRT-LLM, and llama.cpp. Hosted API access is provided by platforms such as OpenRouter, Groq, and Together AI. OpenRouter pricing is set at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens. Fine-tuning is supported through frameworks like TorchTitan and NVIDIA NeMo AutoModel.
