Muse Glimmer 30B is Meta's open-weight local agent model for developers who want a capable AI agent that runs entirely on a single consumer GPU instead of in the cloud. Released on August 10, 2026 under the Apache 2.0 license, it is a 29.6-billion-parameter dense model distilled from the closed Muse Spark 1.2 teacher.
Core Features
- Local-first inference: a 4-bit quantized build fits under 20 GB, running on a 24 GB RTX 4090 or 5090 or an M4 or M5 Max MacBook with 32 GB of unified memory.
- Long context: a 131,072-token window with multimodal text-and-image input across more than 100 languages.
- Agentic tool use: native function calling, multi-step tool use, and LLM-as-judge evaluation.
- Failure recovery: when a tool call fails or returns garbage, Glimmer diagnoses the error and retries instead of stopping.
- Day-zero runtime support: Ollama, LM Studio, llama.cpp, MLX, vLLM, and SGLang, plus hosted inference on Together AI, Fireworks AI, NVIDIA NIM, and OpenRouter.
Use Cases / Best For
- Always-on personal agents that must stay on-device for privacy.
- Local coding agents and SWE-Bench-style workflows on a 24 to 32 GB GPU.
- Synthetic-data generation and LLM-as-judge scoring where per-token API cost would otherwise dominate.
Pricing
The weights are free under Apache 2.0 with no license fee from Meta. Hosted API pricing is set by third parties; Artificial Analysis tracked a median of about $0.32 per million input tokens and $1.35 per million output tokens as of August 17, 2026 on OpenRouter.
Our Take
Best for engineers who need a frontier-adjacent agent that never leaves their laptop; the trade-off is that Qwen3.6-27B still leads on harder coding and computer-use benchmarks, so reach for Muse Spark 1.2 when you need frontier quality. Compare it in our AI Models directory.
