Meta has released Muse Spark Open, a massive open-weight multimodal model specifically trained for agentic coding and complex tool execution. Featuring approximately 816 billion total parameters and a one-million token context window, the model represents the most capable open-weights release from a US company to date.
Muse Spark Open is built to process text, images, audio, and video simultaneously, outputting text. Unlike models trained primarily for conversational chat, Meta optimized this release for long-horizon developer workflows: reading large codebases, writing and debugging scripts, and browsing the web to self-correct during complex tasks.
The Architecture: Massive Scale, Faster Inference
Operating an 816-billion parameter model traditionally requires an impractical amount of compute for most applications. Muse Spark Open addresses this using a Mixture-of-Experts (MoE) architecture.
Rather than running every parameter for every step, the model contains 256 "experts" per layer. A routing mechanism selects only 8 experts for each token. As a result, while the model requires the storage memory of an 816B model—roughly 1.5 TB in its standard BF16 format—it only uses about 42 billion active parameters per token. This allows it to run at the much faster speed of a 42B model.
For deployment, this still requires serious hardware. The standard BF16 checkpoint spans across four nodes of eight H200 GPUs. However, Meta has also released an NVFP4 checkpoint that stores parameters in 4-bit precision for NVIDIA's new Blackwell GPUs, allowing it to run on a single B200 node with negligible accuracy loss.
Multimodal Inputs for Real-World Workflows
The most significant feature of Muse Spark Open is its perception and audio encoders, which allow it to natively process non-text inputs alongside code.
The perception encoder splits images and video frames into patches, converting them into tokens for the language model. Similarly, a dedicated audio encoder converts audio into tokens at a rate of 12.5 tokens per second. Because the model supports a one-million token context window, all of these modalities can be loaded into the same prompt.
In practice, this means the model can read a screenshot of a user interface, listen to a voice note explaining a bug, or watch a screen recording of a software workflow while it actively writes or debugs the code associated with those inputs.
The Bottom Line for SMB Automation
For small and mid-sized businesses automating their operations, the release of a frontier-level open-weight model changes the backend economics and capabilities of custom automation.
First, building reliable automations historically requires perfectly documented standard operating procedures. Muse Spark Open’s ability to ingest video and audio means the friction of translating a human workflow into an automated script drops significantly. A developer can feed the model a screen recording of an employee performing a repetitive software task, and the model can directly generate the agentic code required to replicate it.
Second, the "agentic" focus of the model—its ability to plan, act, observe, and self-correct over long debugging sessions—means automations can become more resilient. Instead of failing when a web interface changes slightly, models trained for this type of long-context self-correction can rewrite their own tool calls on the fly.
While most SMBs will not host a 1.5 TB model on their own servers, the availability of Muse Spark Open through inference providers gives automation partners a powerful, open-source alternative to proprietary models. This ensures businesses can build highly complex, multimodal operations infrastructure without being permanently locked into the ecosystem of a single AI vendor.