Muse Spark is a large language model (LLM) developed by Meta through its Meta Superintelligence Labs (MSL). It is the first model in Meta's Muse family, designed for multimodal reasoning, coding, and AI-assisted software development. The model can handle a million tokens of context and is used across various Meta AI products and services.
Introduced in April 2026, Muse Spark was launched as Muse Spark 1.1 on July 9, 2026. It represents Meta's entry into the competitive landscape of advanced generative AI models, following its earlier Llama series. The model's architecture and training details are proprietary, but it is known to support both natural language and visual inputs, making it a multimodal system.
History and Development
The development of Muse Spark was led by Meta Superintelligence Labs, a division focused on advancing superintelligence research. The model was first announced in April 2026, with the initial release of Muse Spark 1.1 on July 9, 2026. Subsequent updates included Muse Spark 1.2 on August 5, 2026, and Muse Spark 1.3 and 1.3 Max later in the year. The model's rapid iteration reflects Meta's commitment to competing with other major AI developers such as OpenAI, Anthropic, and Google DeepMind.
Muse Spark is built on a transformer architecture, a foundation common to many modern large language models. It incorporates mixture of experts techniques in some variants to optimize performance and efficiency. The model is trained on vast datasets, enabling it to perform tasks such as code generation, debugging, and complex reasoning.
Capabilities and Usage
Muse Spark is designed for multimodal reasoning, meaning it can process and generate both text and visual information. This capability makes it suitable for a wide range of applications, from coding assistance to creative tasks. The model's million-token context window allows it to handle long documents and complex conversations, a feature that distinguishes it from earlier models.
In the realm of AI-assisted software development, Muse Spark can write and debug code, complete longer development tasks, and run multiple sub-agents concurrently. This functionality is leveraged in tools like Muse Code, a terminal-based coding agent released alongside Muse Spark 1.2. The model is also integrated into other Meta AI products, enhancing features such as Meta AI assistants.
Performance benchmarks from the LMArena leaderboard indicate that Muse Spark performs strongly in various categories. As of the latest snapshot, it ranks first in vision humor with a score of 1343.56, fifth in vision creative writing with 1326.99, and ninth in German text with 1498.05. It also ranks 12th in Korean, 13th in vision English, 15th in vision overall, 18th in French, 20th in vision OCR, and 24th in text English. These results highlight its competitive standing in multilingual and multimodal tasks.
Variants
Muse Spark Open
Muse Spark Open, an open-weight version of the model, was released on October 15, 2026. It is a mixture of experts model with 816 billion total parameters and 42 billion active parameters. This variant allows researchers and developers to access the model's weights, facilitating further innovation and customization. The open-weight release is part of Meta's strategy to engage the broader AI community, similar to its earlier Llama releases.
Muse Code
Muse Code is a terminal-based coding agent powered by Muse Spark, released on August 5, 2026, alongside Muse Spark 1.2. It can write and debug code, complete longer software development tasks, and run multiple sub-agents concurrently. Muse Spark 1.2 was made available through the Meta Model API and OpenRouter. A lower-priced contributor tier allows Meta to retain submitted data to improve its products, while data submitted through the higher-priced standard tier is not retained for that purpose.
Muse Glimmer
Muse Glimmer is an open-weight multimodal large language model developed by Meta, released on August 10, 2026. The 30-billion-parameter causal language model combines a large language model with a dedicated perception encoder and supports both natural language and visual inputs. It was designed for local agentic and coding workloads and can be run offline on a 24 GB consumer GPU. The model weights are distributed under the Apache License 2.0, making it accessible for a wide range of applications.
Reception and Impact
Muse Spark has been well-received in the AI community, with its performance on public leaderboards placing it among top-tier models. Its multimodal capabilities and large context window are seen as significant advancements over previous models. The open-weight variants, particularly Muse Spark Open and Muse Glimmer, have been praised for democratizing access to advanced AI technology.
The model's integration into Meta's ecosystem, including Meta AI and other products, has expanded its reach. However, some concerns have been raised about data retention policies in the contributor tier of Muse Spark 1.2, though Meta has stated that standard tier data is not retained for product improvement.
See Also
References
- Koebler, Jason (September 22, 2026). "Meta Tests Muse AI Agent Calls That Are Actually Made By Humans in a Call Center". 404 Media. Retrieved September 28, 2026.
- Benchmark leaderboard snapshot (wikiprompt.org/benchmarks, aggregated daily from public leaderboards).