Muse Glimmer: Meta’s Free Open Model & Price

Muse Glimmer is Meta’s free 30B open agentic model under Apache 2.0. See API pricing, the hardware you need to run it at home and how it stacks up today.

Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, and it’s the first open-weight model from the lab. It’s a 30-billion-parameter dense model distilled from Muse Spark, the model behind Meta Muse, and Meta built it for agents rather than chat: long chains of tool calls, recovering when a call fails, and reading images as well as text. If you searched for “meta muse glimmer”, this is the same model. We checked prices on Together AI and OpenRouter on October 23, 2026.

Is Muse Glimmer free?

Yes. The weights are free to download from Hugging Face under the Apache 2.0 licence, the most permissive licence Meta has used so far. You can use, modify and redistribute it commercially with no user cap. Ollama and LM Studio were launch partners, so you can run it on your own computer and pay nothing per token.

You only pay if you don’t want to host it yourself. Then you use a hosted API, billed per token.

Muse Glimmer Pricing (2026)

There’s no paid plan from Meta. Hosted providers set their own prices:

Option	Input (per 1M)	Cached input	Output (per 1M)
Self-hosted (Ollama, LM Studio, llama.cpp)	$0	$0	$0
Together AI	$0.35	$0.04	$1.50
DeepInfra via OpenRouter (cheapest listed)	$0.30	$0.04	$1.20

At the DeepInfra rate, a job with 1M input and 1M output tokens costs $1.50. That’s more than speed-first models like MiniMax M3.1 Flash, so Glimmer’s real price advantage is that you can run it at home for free. If per-token billing is new to you, read what tokens are in AI first.

What is Muse Glimmer?

• Size: 30B dense parameters, with a perception encoder for image input. Output is text.
• Context: 128K tokens by default. OpenRouter lists 131,072 tokens of context and up to 16,384 output tokens.
• Built for agents: Meta trained it on end-to-end agent tasks, reliable tool use, multi-step reasoning and failure recovery, with controllable effort.
• Languages: over 100.
• Licence: Apache 2.0.

Independent results are decent for its size. Artificial Analysis, as listed on OpenRouter, gives it 17.5 on its Intelligence Index, 51.7% on Terminal-Bench 2.1 and 83.5% on GPQA. The Ornith team’s comparison table lists Glimmer at 76.0 on SWE-bench Verified.

How to run Muse Glimmer locally

The size you download decides the hardware you need:

Build	Download size	Hardware
Q4_K_M GGUF, text only	About 17 GB	A 24 GB GPU, such as an RTX 4090
Q4_K_M GGUF with vision	About 19 GB	A 24 GB GPU
Dynamic GGUF	About 20 GB	A 32 GB Mac or PC
Full bf16	About 72 GB of VRAM	A data-centre GPU

1. Ollama or LM Studio: search for Muse Glimmer in the model library and pick the quantized build that fits your memory.
2. llama.cpp: use build b10353 or newer. Older builds don’t support the model.
3. Coming from Muse Spark Open? Our guide to running Muse Spark Open locally covers the same tools.

For other models that fit on a single consumer GPU, see our list of the best local AI models.

Limitations

• Text output only: it reads images but can’t create them.
• Not a frontier model: larger open models like Ornith 1.5 397B score much higher on coding benchmarks.
• Hardware: you need a 24 GB GPU or a 32 GB Mac for a smooth local setup.
• Short output: hosted providers cap replies at about 16K tokens.

Verdict

Muse Glimmer is the open model to try if you want an agent that runs on your own machine with no licence worries. The Apache 2.0 licence makes it easy to ship in a product. For harder coding work, compare Ornith 1.5, or use Muse Spark Open if you have bigger hardware.

Muse Glimmer is Meta’s first truly permissive open model: a 30B agent model you can run on one 24 GB GPU and ship commercially.

Run it locally for free, or use DeepInfra or Together AI if you’d rather pay per token than manage hardware.