Muse Glimmer 30B

Always-on agentic model built for long-running tool use under Apache 2.0

About model

Muse Glimmer is the latest open-source model from Meta and the first released by Meta Superintelligence Labs: a 30B dense model released under Apache 2.0, the most permissive open-source license from Meta to date, with no restrictions on commercial use, modification, or redistribution. It is built for always-on agents rather than chat alone, designed to handle many sequential tool calls, recover from failures, and handle multimodal input and reasoning. A perception encoder gives the model native image understanding alongside text, and the dense architecture pairs sliding-window attention with periodic global layers across a 128K+ token context window. Optimized for popular open-source agent scaffolds. Available on Together AI.

‍

Most Permissive License from Meta

Apache 2.0

No restrictions on commercial use, modification, or redistribution

Dense Parameters

30B

Full capability on every token, no expert routing

Context Window

128K+

Sliding-window attention with periodic global layers

Model key capabilities

Always-On Agents: Built to handle many sequential tool calls, recover from failures, and perform multi-step reasoning
Scaffold Compatibility: Optimized for popular open-source agent scaffolds, going from endpoint to working agent quickly
Most Permissive License from Meta: Apache 2.0 with no restrictions on commercial use, modification, or redistribution
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

API usage

cURL
Python
Typescript

Endpoint:

meta-models/Muse-Glimmer-30B

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-models/Muse-Glimmer-30B",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="meta-models/Muse-Glimmer-30B",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'meta-models/Muse-Glimmer-30B',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• Dense causal language model, roughly 30B total parameters including a ~1.8B perception encoder for visual input
• 52 transformer layers with sliding-window attention (2,048 tokens) interleaved with a global attention layer every fourth layer
• Gated attention with grouped-query attention: 32 query heads and 2 KV heads at head dimension 128
• SwiGLU feed-forward network, rotary position embeddings applied on local attention layers, 202K vocabulary with untied embeddings
• 128K+ token context window and extendable; DFlash speculative decoding supported

Training Methodology:
• The first model built by Meta Superintelligence Labs and the next open-source release from Meta
• Trained for always-on agentic behavior: sequential tool calling, failure recovery, and multi-step reasoning
• Optimized for popular open-source agent scaffolds

Performance Characteristics:
• Meta highlights agentic performance on TerminalBench 2.1 and OSWorld-Verified as core proof points for agent productivity and reliability
• Built for end-to-end agentic workflows: research, decision, and action loops without constant human input
‍
Prompting
Together AI API Access:
• Access the model via Together AI APIs using the endpoint meta-models/Muse-Glimmer-30B
• Authenticate using your Together AI API key in request headers
• Supports tool calling through the model's chat template for multi-step agent workflows
• Accepts text and image input with text output across a 128K+ token context
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Always-On Agent Systems:
• Run agents that research, decide, and act across sessions lasting hours or days
• Chain many sequential tool calls with failure recovery built into the model's training

Agentic Development Workflows:
• Drop the model into popular open-source agent scaffolds via the Together endpoint
• Build tool-driven pipelines that hold working context across a 128K+ window
• Prototype and ship commercial agent products freely under Apache 2.0

Multimodal Agent Tasks:
• Ground agent decisions in screenshots, documents, and images through the perception encoder
• Combine visual understanding with sequential tool use in one model
• Automate workflows that mix reading interfaces with taking actions
‍