Meta releases Muse Spark Open weights for a large multimodal model

What Meta released

In its launch announcement, Meta says Muse Spark Open is available through hosted APIs and downloadable weights. The company positions it for coding and agentic work—work involving several reasoning steps. This is a provider announcement, not an independent test.

The model and its practical boundary

The model card lists 816 billion total parameters but 42 billion active for each token, a mixture-of-experts design routing work through part of the model. It documents a one-million-token context and accepts text, images, video and audio while producing text. For self-hosting, Meta lists at least 700 GB of VRAM for NVFP4 or 2 TB for BF16, and recommends safeguards tailored to the intended use.

Inspect the fit before buying hardware

Meta’s developer guide lists serverless and dedicated endpoints specifically so teams can evaluate the model before committing hardware. It also separates Meta’s BF16 and NVFP4 checkpoints from formats supplied by hardware and tooling partners. For an initial comparison, we suggest recording which checkpoint and hosting option you use; this is an editorial suggestion, not a reported test result.

Read benchmark claims with their setup

The separate evaluation methodology says benchmark tasks, inputs, grading and primary metrics were preserved as closely as possible, using official sources when they cover all compared models. It also notes that common settings for third-party models may not show their best provider-optimized performance. For a comparison, we suggest recording the benchmark, harness, model settings and source of each result rather than treating a headline score as universal.