--- title: 'Muse Glimmer' description: "Meta Superintelligence Labs' 30B dense multimodal model, distilled from Muse Spark and released with downloadable Apache 2.0 weights for local agent workflows." canonical_url: 'https://darkfactory.dev/glossary/muse-glimmer' markdown_url: 'https://darkfactory.dev/glossary/muse-glimmer.md' collection: glossary date_published: '2026-10-07T00:00:00-04:00' date_modified: '2026-10-07T00:00:00-04:00' --- # Muse Glimmer ## Definition Muse Glimmer is Meta Superintelligence Labs' 30-billion-parameter dense multimodal model for local agent workflows. It accepts text and images and produces text, including reasoning and proposed tool calls. Meta releases its weights under Apache 2.0 for operators to run on their own infrastructure. An agent can use Glimmer's reasoning and proposed tool calls. Its harness supplies persistent state, executes tools, and enforces account permissions. ## Origin and attribution Meta introduced Glimmer on August 10, 2026, crediting Meta Superintelligence Labs. The announcement describes distillation from Muse Spark, followed by additional training for reasoning, coding, and agent tasks. Distillation gives Glimmer a documented training lineage; it does not make its behavior equivalent to Spark. The launch names compatibility with OpenClaw and other scaffolds. These systems supply the surrounding tools and execution controls for using the model as an agent. ## Capabilities and serving scope The specification reviewed October 7 describes a dense decoder-only architecture with a vision encoder and a default 128K-token context window, with support for longer contexts. Verify the context configuration accepted by the chosen runtime. Meta trained Glimmer for tool calling, longer tasks, and recovery after failed tool calls. These are learned behaviors to evaluate with the tools actually exposed. A generated retry can remain wrong, and a valid tool-call schema does not establish a permitted or correct action. Quantized weights reduce memory demands for local inference. A weight-file size does not include the entire serving process: context caches, vision processing, and any drafting model also consume memory. Test the hardware, quantization, context length, and concurrency together before relying on a general claim that it fits one device. ## License, privacy, and limits Apache 2.0 describes the published weights' access and use terms. It does not determine the privacy or retention policy of a service hosting those weights. Self-hosting puts inference under the operator's control; a remote provider running Glimmer has its own deployment terms. Local inference can avoid sending prompts to a hosted model endpoint. Connected email, browsers, external tools, and telemetry can still transmit information. Trace the whole agent's data flow before describing a deployment as private or offline. Provider benchmarks concern selected tasks and serving conditions. Quantization choices and the harness can change the quality measured in practice. Record the model artifact and runtime configuration, and verify results beyond the model's account of what it did. ## Distinguish it from nearby terms - Muse Spark is the larger teacher family in Glimmer's documented lineage, with its own hosted API versions and terms. - Muse names a personal-agent application. Glimmer weights do not include that product's cloud computer, memory interface, or permission service. - OpenClaw is a personal-agent system that can supply a harness around a model. Selecting Glimmer does not set OpenClaw's permissions. - Quantization changes how model parameters are represented for serving. It does not grant the agent authority or guarantee an unchanged result on every task. ## Check your understanding A local Glimmer agent reads private files and uses a hosted search tool. Which paths keep information on the device, and which can send it elsewhere? What evidence would support calling the complete deployment offline? ## Also called Muse Glimmer model ## Related terms - [Muse models](https://darkfactory.dev/glossary/muse-model) - [Muse Spark](https://darkfactory.dev/glossary/muse-spark) - [Muse](https://darkfactory.dev/glossary/muse) - [OpenClaw](https://darkfactory.dev/glossary/openclaw) - [Personal agent](https://darkfactory.dev/glossary/personal-agent) - [AI model](https://darkfactory.dev/glossary/ai-model) - [Large language model (LLM)](https://darkfactory.dev/glossary/large-language-model) - [Multimodal model](https://darkfactory.dev/glossary/multimodal-model) - [Reasoning model](https://darkfactory.dev/glossary/reasoning-model) - [Open-weight model](https://darkfactory.dev/glossary/open-weight-model) - [Model distillation](https://darkfactory.dev/glossary/distillation) - [Quantization](https://darkfactory.dev/glossary/quantization) - [Context window](https://darkfactory.dev/glossary/context-window) - [Model routing](https://darkfactory.dev/glossary/model-routing) - [Agent harness](https://darkfactory.dev/glossary/agent-harness) - [Evaluation (eval)](https://darkfactory.dev/glossary/evaluation) ## Related factory areas - [Model selection, routing & budgets](https://darkfactory.dev/factory/model-routing-budgets) - [Execution environments, identity & secrets](https://darkfactory.dev/factory/execution-environments) - [Context, memory, knowledge & skills](https://darkfactory.dev/factory/context-memory-skills) - [Verification, evaluation & quality truth](https://darkfactory.dev/factory/verification) - [Security, privacy & software supply chain](https://darkfactory.dev/factory/security) ## Evidence and further reading - [Meta: Introducing Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) - [Meta: Muse Glimmer specification](https://dev.meta.ai/docs/muse-glimmer) - [Meta: Model catalog](https://dev.meta.ai/docs/models)