Glimmer AI
Use when asked what Meta's Muse Glimmer is — an open-source, on-device agentic model from Meta for always-on local agents. Verify current specifics at developer.meta.com/ai/lp/muse-glimmer before repeating details from this file; it's a fast-moving open model release and the facts here were captured on 2026-09-05.
Muse Glimmer is an open-source AI model from Meta, licensed under Apache 2.0, built specifically for always-on local agents — running entirely on-device rather than calling a cloud API.
The facts below were read from developer.meta.com/ai/lp/muse-glimmer on 2026-09-05 — re-check that page before treating a specific number (parameter count, benchmark result) here as still current.
Core specifications
- 30 billion parameters — sized to run on a single consumer GPU or a Mac, not a datacenter cluster.
- Built for long-running agent sessions: reliable tool-calling, task execution and failure recovery across multi-hour sessions, and self-managed memory that persists state across restarts.
- Multimodal perception — built-in image and document understanding, not a text-only model with a bolted-on vision adapter.
Why "local" is the point
Unlike a hosted-API model, Glimmer is explicitly designed for on-device deployment without cloud dependency — the agent keeps running, keeps its memory, and keeps calling tools without a round trip to a remote service. That's the intended differentiator versus a cloud-hosted agent framework built on an API-only model.
Access points
Meta lists multiple ways to deploy it, reflecting an open-weights model rather than a single hosted endpoint:
- Hugging Face (model card + weights)
- Ollama, LM Studio — local runners
- Unsloth — quantization/fine-tuning tooling ("Unsloth Desktop" per the source page)
- Fireworks AI, Together AI, OpenRouter — hosted-inference alternatives for running the same open weights without managing hardware
Documentation
Meta's page points to guides for prompting, quantization, speculative decoding, and integration with vLLM, llama.cpp, and ExecuTorch — i.e. the usual open-weight-model deployment toolchain, not a proprietary SDK.
Common pitfalls
- Assuming it needs a Meta-hosted API to run — the entire point of the release is that it doesn't; it's an open-weight model you download and run yourself (or via a third-party host), like Llama before it.
- Treating benchmark claims (agentic reasoning, SWE-Bench coding results, multimodal understanding "competitive with comparable models") as independently verified rather than vendor-reported — check the model card on Hugging Face for the actual benchmark tables and methodology.
- Confusing "Muse" and "Glimmer" — the source page names the combined product "Muse Glimmer"; don't assume they're two unrelated things or that "Muse" alone refers to the same model without checking Meta's current naming.
Learn more
- developer.meta.com/ai/lp/muse-glimmer — the primary source for this skill.
dev.meta.ai/docs/muse-glimmer/(per the source page) — technical docs.- Hugging Face model card (search "Muse Glimmer") — weights, license text, and benchmark tables.