meta-models/Muse-Glimmer-30B

Hugging Face
VISIONPricing:Input $0.3 / Cached $0.04 / Output $1.1Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Aug 9, 2026License:apache-2.0Architecture:Transformer1.9K Open Weights Warm

Muse Glimmer is a 30-billion-parameter causal language model developed by Meta Superintelligence Lab, featuring a dedicated perception encoder for multimodal understanding. Optimized for autonomous agentic tasks, it integrates multi-step reasoning, reliable tool use, and failure recovery, designed to run efficiently on consumer hardware. The model excels at end-to-end agentic task completion, coding, and multimodal reasoning, supporting local deployments with quantized versions for 24GB/32GB VRAM. It is multilingual, trained on over 100 languages, and includes speculative decoding for faster generation.

Loading preview...

Muse Glimmer-30B: Autonomous Agentic AI for Local Deployment

Muse Glimmer is a 30-billion-parameter causal language model from Meta Superintelligence Lab, specifically engineered for autonomous agentic tasks on consumer hardware. It integrates a dedicated perception encoder, enabling multimodal understanding by accepting interleaved text and images. The model is designed for local execution, offering capabilities like multi-step reasoning, reliable tool use, and robust failure recovery without cloud infrastructure.

Key Capabilities & Features

  • End-to-end Agentic Task Completion: Achieves strong success rates on benchmarks like DeepSearch QA, MCP-Atlas, and SWE-Bench, demonstrating proficiency in code writing, debugging, and multi-turn request resolution.
  • Multimodal Understanding: Processes text and images, allowing agents to interpret screenshots, charts, and documents.
  • Optimized for Local Deployment: Utilizes 4-bit quantization to fit within 24GB or 32GB VRAM, with minimal degradation on agentic tasks. Includes a DFlash speculative decoding drafter for significantly faster text generation (e.g., 3.1x speedup on Nvidia RTX 5090).
  • Reliable Tool Use & Failure Recovery: Handles diverse function calls with precise schemas and diagnoses/retries failed tool calls.
  • Multilingual Support: Trained on data from over 100 languages.
  • Controllable Effort: Supports varying reasoning strengths (low/medium/high/xhigh) for balancing quality and speed.

Ideal Use Cases

  • Local AI Agents: For multi-step planning, sequential tool invocation, and long-horizon task execution on consumer devices.
  • Coding Agents: Excels at writing, debugging, and resolving software engineering tasks.
  • Multimodal Reasoning: Interpreting visual information alongside conversational context.
  • Synthetic Data Generation & LLM-as-a-Judge: For developing and evaluating other models.