DataAIHub
DataAIHubNews · Research · Tools · Learning
MetaOpen SourceSmall ModelsMultimodalVision

Muse Glimmer

Meta’s Apache-2.0 ~30B open multimodal agent model for on-device and single-GPU local agents.

Meta Superintelligence Labs’ Muse Glimmer — Apache-2.0 ~30B dense multimodal agent model for on-device and single-GPU local agents. Sibling to closed Muse Spark; distinct from Llama 4.

Why Muse Glimmer matters

Glimmer is Meta’s open Muse checkpoint for local agents—not Llama 4 and not closed Muse Spark. Use it when you want Apache-2.0 weights, consumer-GPU deployability, and agentic/tool loops without a Meta API. Escalate to Spark/Llama/frontier APIs when you need 1M context, hosted SLA, or harder open MoE capability.

Vision · Tool calling · Thinking · MCP · Coding

Last reviewed: 25 August 2026

Official pricing →

When to choose Muse Glimmer

Decision guidance for architects—not a feature list.

Best for

  • On-device / single-GPU personal agents
  • Privacy-sensitive local coding and tool use
  • Offline multimodal assistants (image+text)
  • Local LLM-as-judge and eval harnesses

Avoid if

  • You need frontier open MoE scale (prefer Kimi K3 / DeepSeek V4)
  • You require Meta-hosted API SLAs (prefer Muse Spark)
  • You want the broadest Llama Hub fine-tune ecosystem

Strengths

Qualitative snapshot for architects—not a public ranking.

  • License (Apache-2.0)★★★★★
  • On-device / single-GPU fit★★★★★
  • Agentic mid-size quality★★★★☆
  • Frontier open MoE scale★★☆☆☆
  • Hosted Meta API★☆☆☆☆

Ecosystem

Built by

  • Meta

    Muse Glimmer is Meta Superintelligence Labs’ open on-device agent model.

Competes with

  • Qwen3

    Competes in the open mid-size / local VLM and agent class (e.g. Qwen3.8-27B).

  • Phi

    Competes for on-device and constrained local assistants; Glimmer is larger (~30B) and more agent-oriented.

Works with

  • vLLM

    Common production serving path for open Muse Glimmer weights.

  • Ollama

    Local runner path as community and Meta partner support lands.

Recommended for

  • Llama Models

    Clarify Llama vs Muse Spark vs Muse Glimmer when evaluating Meta stacks.

  • AI Agents

    Designed for local agent scaffolds, tool use, and always-on personal agents.

Often paired with

  • Muse Spark

    Spark is the closed Meta Model API / Muse Code path; Glimmer is the Apache-2.0 local weights path.

  • Llama 4

    Llama remains Meta’s general open-weight family; Glimmer is the Muse-line local agent checkpoint.

How Muse Glimmer evolved

Key moments in chronological order.

  1. Open source

    Muse Glimmer 30B open weights

    Apache-2.0 ~30B dense multimodal agent model for on-device / single-GPU use; Hugging Face meta-models/Muse-Glimmer-30B. Distinct from closed Muse Spark and from Llama 4.

  2. Platform

    Weights + quantizations on Hugging Face

    BF16 full weights, 4-bit variants for 24/32 GB GPUs, perception encoder, and DFlash drafter head published under Apache-2.0.

  3. Partnership

    Edge and serving partner path

    Meta points developers to llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, vLLM, SGLang, and hosts like Together / Fireworks / OpenRouter as integrations land.

  4. Model

    Muse Spark 1.2 sibling context

    Closed Muse Spark 1.2 / Muse Code beta ship days earlier as the paid Meta Model API path; Glimmer is the open local counterpart.

Overview

Meta Superintelligence Labs’ Muse Glimmer — Apache-2.0 ~30B dense multimodal agent model for on-device and single-GPU local agents. Sibling to closed Muse Spark; distinct from Llama 4.

Capabilities

  • Vision: Yes
  • Audio: No
  • Tool calling: Yes
  • Thinking: Yes
  • MCP: Yes
  • Coding: Yes
  • Structured output: Yes

Technical specifications

Provider
Meta
License
Apache-2.0
Context window
131K
Parameters
~29.6B dense (+ ~1.8B vision encoder)
Architecture
Dense Transformer + Perception Encoder
Release
2026-08
Modalities
Text, Image
Vision
Yes
Audio
No
Tool calling
Yes
Thinking
Yes
MCP
Yes
Open weights
Yes
API
No
Pricing (input)
Self-host / third-party hosts (no Meta API for Glimmer at launch)
Pricing (output)
Self-host / third-party hosts

Supported modalities

Text · Image

Context window

131K (131,072 tokens)

Pricing

Input: Self-host / third-party hosts (no Meta API for Glimmer at launch)
Output: Self-host / third-party hosts

Weights on Hugging Face under Apache-2.0. Hosted pricing depends on providers (Together, Fireworks, OpenRouter, etc.).

Availability

API: No
Chat UI: No
Open weights: Yes

Download meta-models/Muse-Glimmer-30B (BF16 + 4-bit). Local: llama.cpp / MLX / ExecuTorch / Ollama / LM Studio as support lands. Serve with vLLM / SGLang.

Use cases

  • On-device / single-GPU personal agents
  • Local coding and tool-calling loops
  • Privacy-sensitive multimodal assistants
  • LLM-as-judge and offline eval harnesses

Strengths

  • Apache-2.0 (more permissive than Llama Community License)
  • Sized for one consumer GPU (24–32 GB class with quantization)
  • Strong agentic mid-size results vs Gemma/Qwen peers per Meta evals
  • Native image+text input for local multimodal agents

Limitations

  • Not a frontier MoE — below Kimi K3 / DeepSeek V4 on hardest open tasks
  • No Meta-hosted API at launch (self-host or third-party only)
  • Newer ecosystem than Llama Hub fine-tunes and tooling
  • 131K context class — shorter than Spark’s 1M API path

Related guides

Related benchmarks

Related research

Related GitHub

Related tools

Related rankings

Companies

Explore more models

All models →