Meta Opens Muse Glimmer for Local AI Agents
Meta Platforms helped establish the open-weight AI market with its Llama models, but its commitment came under scrutiny as frontier development became more guarded. Muse Glimmer marks a renewed push to put capable models in developers’ hands. Built by Meta Superintelligence Labs and derived from the larger Muse Spark, it targets agents that can reason, write code and use tools while keeping sensitive data on-device and reducing dependence on cloud services.
Meta released Muse Glimmer on Aug. 10, 2026, under the Apache 2.0 license. The 30-billion-parameter model accepts text and images and supports a context window of at least 131,072 tokens. A 17-gigabyte quantized build can run on a single consumer GPU with 24 GB of memory. Meta said that version trails the full-precision model by only about 1% on average across 15 benchmarks, while its DFlash speculative-decoding system is designed to accelerate the repeated inference calls required by agent workflows.
All Coverage
7 original reportsThe Backstory
The history behind this eventMeta’s Muse Spark 1.2 Posts Gold-Level Reasoning, Strong Benchmark Gains
Meta Superintelligence Labs introduced the Muse Spark family in April as a multimodal reasoning system designed for complex scientific and software tasks. Muse Spark 1.2 extends that push as leading AI developers increasingly compete not only on headline accuracy, but also on the cost and speed of deploying models. Its tool-free Olympiad tests are significant because they seek to isolate the model’s underlying reasoning rather than gains from web search, coding or calculators.
Meta released Muse Spark 1.2 on Aug. 5 and on Aug. 7 said it achieved gold-medal-level results across several STEM Olympiads without external tools. Vals AI scored the model at 71.9 on the Vals Index, up 3.5 points from Muse Spark 1.1, placing it in the top five at a cost of $0.69 per test. It also led Finance Agent v2 with 59.4%, ahead of Gemini 3.6 Flash at 58.1%, while completing the Index run in 629.7 seconds.
Meta Launches Multimodal AI Model Muse Spark Across Social Platforms and Smart Glasses
Meta built its open-source AI profile with the Llama series, but is now introducing Muse Spark from Meta Superintelligence Labs, a multimodal model designed to process text and images. The model will be integrated into Facebook, Instagram, WhatsApp and smart glasses, signaling a shift toward a product-first strategy and commercial applications.
Muse Spark was formally unveiled after three delays. As of July 19, 2026, its API was available only to invited developers in preview and the model had not been open-sourced. Meta has not yet disclosed API pricing, model parameters or the amount invested, but has confirmed that Muse Spark will strengthen Meta AI’s reasoning capabilities and extend into personal-assistant services for areas including health and shopping.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.