Muse Spark Open unveiled: A multimodal model with 816 billion parameters

Meta’s open-weight multimodal model Muse Spark Open was introduced on 15 October 2026 in a Hugging Face blog post. Described as being trained for coding workflows, the model can process text, image, audio and video inputs and generate text. It has approximately 816 billion total parameters, and its Mixture-of-Experts (MoE) architecture activates about 42 billion parameters for each token.

What inputs can Muse Spark Open process?

The model can work with different media types in a single system. Images and video frames are passed to the text model through a perception encoder, while audio is passed through a separate audio encoder. Hugging Face’s post gives examples of the model reading screenshots, listening to voice notes and reviewing screen recordings. The model’s output is limited to text.

By default, video is sampled at 2 frames per second, with up to 96 frames processed. The frame rate can be increased, but the source notes that doing so may raise memory requirements and preprocessing latency. When audio is provided directly, the expected format is mono at a 24 kHz sample rate. The model’s context window can reach 1 million tokens, counting text and media tokens together.

MoE architecture and hardware requirements

Each MoE layer in Muse Spark Open has 256 experts, of which 8 are selected for each token. This design keeps the number of active parameters during computation at approximately 42 billion; however, all experts must be held in memory. As a result, the number of active parameters alone does not indicate how much memory is needed to run the full model.

According to the technical details shared by Hugging Face, BF16 weight files take up approximately 1.5 TB in safetensors format. Running the BF16 version with Transformers requires 4 nodes with 8 H200 GPUs, while the NVFP4 version requires a single B200 node. NVFP4 is described as a 4-bit format designed for NVIDIA Blackwell GPUs. The post also estimates that a full 1-million-token context could require approximately 266 GB of KV cache in BF16.

Coding and deployment options

The model was trained with a focus on developer workflows such as multi-file refactoring and long debugging sessions. It also supports tool calling, and can approach long tasks by planning, taking action, observing results and self-correcting. These capabilities suggest that the product’s primary focus is not direct image generation, but coding and tool use with multimodal inputs.

The Hugging Face post identifies the MuseSparkForConditionalGeneration and MuseSparkProcessor components in Transformers. DistributedConfig is used for distributed execution settings. The post also links to run guides for llama.cpp, vLLM and SGLang. The shared text does not specify pricing or licensing details; the license and hardware requirements in the model repository should be checked separately before deployment.

Implications for architecture and visualization workflows

For architects, interior designers and visualization teams, the model’s practical appeal may lie in its ability to review images and video recordings alongside text. For example, asking questions about a screenshot or a recording of a work session could be tested in design and technical review processes. However, the source does not announce architectural software plugins or a dedicated render-analysis feature. This is therefore a potential use arising from multimodal input support, not an existing stated capability.

The main hurdle is hardware. The configurations described require systems with very high memory capacity, so offices considering local use should not look only at the number of active parameters per token. Weight format, GPU compatibility, memory requirements and licensing conditions should all be assessed together.