Meta Muse Glimmer
Category: Foundation Models / LLMs
An open-weight 30B-parameter multimodal agentic model released by Meta Superintelligence Labs under Apache 2.0, optimized to run autonomous AI agents locally on a single consumer GPU. The company is led by Mark Zuckerberg. Based in Menlo Park, California, USA.
- Headquarters
- Menlo Park, California, USA
Value proposition
Muse Glimmer is an open (Apache 2.0) 30B-parameter multimodal agentic model designed for always-on local agent workflows. It runs on a single consumer GPU (24GB/32GB) via 4-bit quantization, enabling local agents, function calling, coding, and LLM-as-a-judge evaluation without cloud/network dependency. It delivers competitive agentic and coding performance for its size class versus Gemma4-31B and Qwen3.6-27B.
Products and solutions
Muse Glimmer-30B (open-weight multimodal agentic model, Apache 2.0), Muse Glimmer-30B-assistant variant, DFlash block-diffusion speculative decoding drafter, quantized K-Quant-17GB model, Muse Spark (larger closed teacher model), Muse Image (media generation).
Unique value
Meta's first Apache 2.0-licensed open-weight agentic model — a 30B model distilled from the much larger Muse Spark teacher that fits on a single consumer GPU, combining multimodal perception, long-horizon agentic reasoning, failure recovery, and speculative decoding for fast local inference.
Target customer
Developers and enterprises building local, on-device autonomous AI agents; open-source community; developers needing local coding, function calling, and LLM-as-a-judge evaluation.
Industries served
AI developer tools, autonomous agents, local/edge AI, coding assistants, multimodal AI applications.
Technology advantage
Compact architecture with hybrid attention (grouped-query attention GQA + sliding window attention SWA at a 3:1 local:global ratio); logit distillation from Muse Spark teacher; ~4-bit quantization shrinking the 30B model to under 20GB to fit a 24GB/32GB GPU; DFlash block-diffusion speculative decoder that proposes whole token blocks for faster generation; dedicated perception encoder for interleaved text+image input; trained on 100+ languages; evaluated under Meta's Advanced AI Scaling Framework.
How they differentiate
Differentiates via Apache 2.0 permissive licensing (Meta's first open-weight agentic model), on-device deployment on a single consumer GPU, hybrid GQA/SWA attention, DFlash block-diffusion speculative decoding for speed, and distillation from the much larger closed Muse Spark teacher model.
Main competitors
Gemma4-31B (Google), Qwen3.6-27B (Alibaba), other open-weight local agentic models
Key partnerships
Optimization partners: AMD, Arm, Dell, Intel, NVIDIA, deployment partners: Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, OpenRouter, edge frameworks: llama.cpp
Major milestones
Released August 10, 2026 as Meta's first Apache 2.0-licensed open-weight agentic model, distilled from Muse Spark (Meta's flagship multimodal reasoning model launched April 2026), runs on a single consumer GPU (MacBook M4/M5 Max, RTX 5090), evaluated on DeepSearch QA, MCP-Atlas, tau-Bench, and SWE-Bench.
Market positioning
Positioned as Meta's return to open-source leadership in agentic AI — an Apache 2.0-licensed 30B model that undercuts cloud-dependent frontier models by running locally on consumer hardware, competing directly with Google's Gemma and Alibaba's Qwen in the open-weight local-agent segment.
Geographic focus
Global (open-source AI market)
Patents and IP
Apache 2.0 open-source license; open weights on Hugging Face (meta-models/Muse-Glimmer-30B).
About Mark Zuckerberg
Co-founder, Chairman and CEO of Meta Platforms (formerly Facebook); Harvard University (dropped out). Founded Facebook in 2004.
Latest news about Meta Muse Glimmer
More Foundation Models / LLMs companies
Official website: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
