Announces Meta AI releases including the Muse image model family, SAM 3, and open-source models accessible via a new Model API.
Meta AI has announced a suite of new models and tools, including the Muse Image family, SAM 3, and open-source models accessible via the Meta Model API. The flagship Muse Spark series now includes Muse Spark 1.3, a proprietary agentic coding model released on September 2, 2026, which offers improved efficiency with roughly 20% fewer tool calls and 25% fewer tokens than its predecessor.
For local deployment, Meta released Muse Glimmer, a 30-billion-parameter open-weight model under the Apache 2.0 license. Distilled from Muse Spark, Glimmer is optimized for coding and agentic tasks, capable of running offline on a single consumer GPU. While Muse Spark remains a closed, commercially accessed model, Meta has indicated plans to release the weights for Muse Spark 1.2 in the near future.
The material is an announcement of new Meta AI research releases, centered on vision models and generative image capabilities. Its headline contribution is the Muse family—Muse Spark, Muse Glimmer, and Muse Image—presented as a suite of image models rather than a single monolithic system. Alongside this, the page highlights SAM 3, an update to Meta’s Segment Anything line, which advances promptable segmentation and related visual grounding tasks. A notable practical component is the introduction of a Model API and open-source access, positioning these releases as usable building blocks for researchers and developers rather than purely academic artifacts.
Technically, the significance lies in the combination of improved vision models, segmentation capabilities, and deployment-oriented access. SAM 3 matters because segmentation is a foundational primitive for editing, measurement, robotics, AR/VR, and downstream generative workflows; stronger promptable segmentation can reduce annotation costs and improve controllability in visual pipelines. The Muse models matter because they expand Meta’s contribution to the image-generation and image-understanding stack, potentially offering open alternatives to closed systems and enabling researchers to study prompt-following, visual fidelity, and controllability under more transparent conditions.
More broadly, the material matters because it lowers the barrier to using advanced vision models through a standardized API and open-source availability, where applicable. For practitioners, this supports faster prototyping, benchmarking, and integration into production systems; for the research community, it improves reproducibility and enables derivative work. The announcement is therefore best read as a platform-level update: Meta is not only releasing new models, but also expanding the infrastructure for accessing and building on them.