Meta has released Muse Glimmer, a 30-billion-parameter open-weight agentic model under Apache 2.0 license, capable of running on a single consumer GPU. Announced August 10, 2026, it’s the first model from Meta Superintelligence Labs to ship with full weights publicly available—and the clearest signal yet of where the company draws the line between closed and open AI.
Technical achievement
A 30B parameter model typically requires over 55 GB of memory at full precision—beyond any consumer GPU. Meta solved this through aggressive quantization to roughly 4-bit, combined with block-wise compression techniques that shrink the model to fit in a single 24GB GPU while maintaining reasonable performance.
The model delivers approximately 20,000 tokens per second on NVIDIA consumer hardware, according to Meta’s benchmarks, with a context window exceeding 120,000 tokens. That’s sufficient for always-on agent workflows that process documents, codebases, and communications locally.
Distilled from frontier
Glimmer is distilled from Muse Spark, the closed frontier model Meta launched in April. It inherits agentic capabilities—function calling, tool use, multi-step reasoning—while being small enough to run without cloud connectivity.
This is the second Muse model released in five months, but the strategy differs from previous releases. April’s Muse Spark stayed closed. July’s Muse Spark 1.1 shipped only as a hosted API. Glimmer lands on Hugging Face with full weights under Apache 2.0.
The local agent thesis
The release reflects Meta’s bet that the future of agentic AI belongs not to the cheapest API, but to agents that run persistently on personal devices. No latency, no cloud bills, no data leaving your machine.
For developers, Glimmer offers something unavailable from OpenAI or Anthropic: a production-grade agent model you can modify, fine-tune, and run privately without API costs.
Why it matters
Muse Glimmer demonstrates that frontier-class agent capabilities can fit in consumer hardware. If the model performs as advertised, it could trigger a shift toward local AI assistants that handle scheduling, coding, and workflow automation without cloud round-trips.
Meta’s choice of Apache 2.0—rather than a more restrictive license—ensures the model can be commercialized by anyone. The company’s AI strategy now spans both closed frontier models (Muse Spark) and open local agents (Glimmer), covering both cloud and edge.