Meta Launches Muse Spark 1.2 and Muse Glimmer AI Models

On August 5, 2026, Meta launched Muse Spark 1.2, a flagship coding model tailored for complex software engineering tasks. This model features a one-million token context length, enhanced first-attempt coding accuracy, and deep integration with persistent asynchronous background agents. Muse Spark 1.2 is specifically designed for debugging extensive codebases, validating changes across numerous files, and managing multi-step reasoning tasks. The significant one-million token context window allows the model to retain the entire scope of a project in its working memory, rather than relying on isolated code snippets.

In conjunction with this release, Meta introduced Muse Glimmer on August 10, 2026. This 30-billion-parameter AI model is optimized to run on a single consumer GPU, making it accessible for users without the need for extensive cloud infrastructure. Muse Glimmer, a distilled version of Muse Spark 1.2, utilizes techniques such as logit distillation and quantization to achieve a compact size of under 20 GB while maintaining most of its reasoning capabilities.

Muse Glimmer is designed for agentic tasks, including schedule management, file organization, coding assistance, and multimodal reasoning, and supports over 100 languages for global deployment. Unlike previous models, Muse Glimmer is open source, with its weights published under the Apache 2.0 license on Hugging Face, allowing for integration with popular inference frameworks.

Meta's shift from an open weights strategy to a closed, API-centric model for its Muse lineup marks a significant change in how the company approaches the commercialization of its AI developments. Muse Spark 1.2 is available exclusively through the Meta Model API, with pricing starting at $1.25 per million input tokens.