Muse Glimmer
Meta's return to open weights — a 30B dense multimodal agent model under Apache 2.0, distilled from Muse Spark to run always-on local agents on a single consumer GPU
Muse Glimmer
Meta • August 2026
Training Data
Distilled from Muse Spark's data mix, plus longer-context agent data with reasoning traces; 100+ languages
Muse Glimmer
August 2026
Parameters
~30 billion (dense)
Training Method
Logit distillation from Muse Spark, then SFT with on-policy distillation and RL
Context Window
131,072+ tokens
Knowledge Cutoff
January 2026
Key Features
Open Weights (Apache 2.0) • Local Agents on Consumer GPUs • Multimodal Perception Encoder • Tool Use with Failure Recovery
Capabilities
Agentic Tasks: Excellent (MCP Atlas 75.5)
Coding: Excellent (SWE-Bench Verified 76.0)
Reasoning: Excellent (AIME 94.7, GPQA 83.5)
What's New in This Version
First open-weight release from Meta Superintelligence Labs, ending the proprietary-only Muse era — quantized checkpoints under 20GB run on 24GB consumer hardware, released alongside Zuckerberg's 'The Future is for Everyone' manifesto
Meta's return to open weights — a 30B dense multimodal agent model under Apache 2.0, distilled from Muse Spark to run always-on local agents on a single consumer GPU
What's New in This Version
First open-weight release from Meta Superintelligence Labs, ending the proprietary-only Muse era — quantized checkpoints under 20GB run on 24GB consumer hardware, released alongside Zuckerberg's 'The Future is for Everyone' manifesto
Technical Specifications
Key Features
Capabilities
Other Meta Models
Explore more models from Meta
Muse Spark 1.1
Meta Superintelligence Labs' strongest model for agentic and coding work — the first Muse model developers can build on, via the new Meta Model API
Muse Spark
Meta Superintelligence Labs' first model — a natively multimodal reasoning model with contemplating mode and multi-agent orchestration
Llama 4 Behemoth
Meta's flagship multimodal model (RESEARCH PREVIEW ONLY - weights not publicly released)
