Meta Platforms, Inc. (NASDAQ: META) has released Muse Spark Open, an open-weight version of its Muse Spark model designed for long-horizon agentic work such as coding and customer-service agents. The company said the multimodal model can accept text, image, video and audio inputs and supports a context window of up to one million tokens.
Meta Muse Spark Open is an 816-billion-parameter sparse mixture-of-experts model that activates about 42 billion parameters per token, according to the company. The architecture is intended to make it possible for an agent to work across a large codebase, plan a change, call tools, check the result and correct a failed step. Meta released the weights under its Muse Model License rather than limiting access to a hosted API.
For enterprise AI teams, the practical significance is deployment choice. Meta said the model is available at launch across several serving runtimes and can be run through hosted inference providers as well as deployed by an organization itself. It also listed managed model-catalog access and options for dedicated endpoints, providing teams ways to evaluate the model before committing to infrastructure.
That flexibility comes with substantial capacity requirements. Meta said the full-precision BF16 checkpoint is about 1.6 TB, while its NVFP4 checkpoint is about 530 GB and can be served on four Blackwell Ultra GPUs or eight H200 GPUs. A quantized local version from Unsloth can run in roughly 300 GB of combined system RAM and GPU memory, according to the source. Those specifications make Muse Spark Open a fit for well-resourced workstation or data-center environments rather than a typical developer laptop.
Meta also described multiple ways to customize the model. It said supervised fine-tuning, LoRA and reinforcement-learning workflows are available at launch through open-source frameworks and managed APIs. Organizations that want to alter behavior for an internal workflow can therefore choose between operating training infrastructure themselves or using a managed post-training service.
Inference optimization is another focus. The release includes draft models for speculative decoding, a technique in which a smaller model proposes several tokens and the larger model validates them. Meta said one supported configuration achieved higher throughput in testing while preserving accuracy, but those figures come from the vendors involved and are not independent benchmarks.
Meta positioned the model for tool calling and multi-step agent loops, areas where enterprises are increasingly testing automation for software development, research and customer operations. The company said the model has support for function calling and structured outputs through several runtimes. The release does not itself establish that any particular enterprise use case will be accurate, secure or cost-effective; teams will still need evaluation, access controls and monitoring before a model is allowed to take consequential actions.
The immediate benefit of Meta Muse Spark Open is the ability to select a deployment path that fits an organization’s data, latency and infrastructure requirements while retaining access to an open-weight model built for agent workflows. Its operational test will be whether the broad runtime, quantization and fine-tuning support lowers the engineering burden of running a large model without shifting too much of that burden into capacity planning, governance and ongoing model operations.