Introducing Muse Glimmer: Meta’s Apache 2.0 Open-Weight Model for Agentic Tasks

Meta has launched Muse Glimmer, a 30-billion-parameter open-weight AI model. Released under a permissive Apache 2.0 license, it’s fine-tuned for finishing complete, agent-driven jobs. The model performs well on benchmarks for code writing, debugging, and logical reasoning.

Key takeaways

• Meta’s Muse Glimmer is a 30B parameter model released under a clean Apache 2.0 license.
• It’s fine-tuned for end-to-end agentic task completion, handling workflows from start to finish.
• The model achieves strong results on benchmarks like SWE-Bench and DeepSearch QA for coding and multi-turn requests.
• Core capabilities include reliable tool use with precise schemas and multi-step reasoning over long horizons.
• Its 30B size is practical for local deployment on machines with 32GB+ of RAM.

Meta’s Return to Open Weights with a Clean License

Meta has released Muse Glimmer, a new 30-billion-parameter model, marking its return to open-weight AI. The model uses a straightforward Apache 2.0 license, a clear upgrade from older, more restrictive licenses.

Meta states the model is fine-tuned for finishing agent-driven jobs from beginning to end. It shows high success rates on benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench, and SWE-Bench. These tests measure its skill in structured environments, code generation, and completing multi-step user requests.

Muse Glimmer also handles tool use reliably. It manages many function calls and uses tools with exact schemas during long jobs. Additionally, it reasons through multiple steps, maintaining logical plans across involved workflows.

For practical use, the 30B size benefits machines with 32 GB of RAM or more. This leaves enough memory for other programs to run at the same time.

Optimized for End-to-End Agentic Task Completion

Muse Glimmer is specifically fine-tuned for finishing full agent-driven jobs. Meta built the model for this exact type of complete workflow execution.

Strong Performance on Full-Task Benchmarks

The model achieves high success rates on benchmarks measuring complete, multi-step jobs. These include DeepSearch QA, MCP-Atlas, 𝛕-Bench, and SWE-Bench. These tests evaluate the model’s ability to operate in frameworks, write and fix code, and solve multi-turn requests entirely.

Core Capabilities: Tool Use and Multi-Step Reasoning

Muse Glimmer demonstrates reliable tool use. It handles many function calls with precise schemas across long workflows. The model’s design for end-to-end agentic jobs includes invoking tools throughout these processes.

Sustaining Coherent Plans

The model excels at multi-step reasoning, linking its logic over extended sequences. This skill lets it keep coherent plans across involved, lengthy workflows. Benchmarks requiring full resolution of multi-turn requests confirm its performance.

Practical Local Deployment and Developer Testing

The Muse Glimmer model’s 30 billion parameter size suits local deployment well. A machine with 32 gigabytes of RAM or more can run it and still have resources for other applications. The author gives their own 128-gigabyte system as an example.

Early Testing with Tool Use

Initial developer tests show the model’s capabilities for agent-driven jobs. One test used the `llm-coding-agent` plugin with the prompt “how does auth work?” on a new Datasette codebase checkout. The model ran a series of tool calls to examine the code and craft a reply, proving its reliable function-calling skill in a long workflow.