At a glance
- Maker: Meta
- Released: 2026
- Parameters: 30B
- Context window: 131k
- Size: Medium
- Memory: Handles long documents
- Good at: Agents on modest hardware
- Supports: tools, zero data retention
- License: Meta's own open licence (community-style terms)
- Upstream API price: $1.50 per million tokens
Announced in Introducing Muse Glimmer: An Open Agentic Model.
What is Muse Glimmer 30B?
Muse Glimmer 30B comes from Meta Superintelligence Labs, and it represents something unusual: a frontier lab deliberately shrinking its best work until it runs on ordinary hardware. The model is distilled — compressed, essentially — from a much larger Meta system, keeping as much of the parent's agentic ability as will fit in 30 billion parameters. The point is not to be the biggest model on the site; it is to be the most capable one you can realistically run yourself.
Its remit is agents: it supports tool use natively, it is covered by zero-data-retention terms, and its 131k-token context window — roughly 100,000 words — is sized for realistic agent loops rather than archive-dumping. It is focused on text and tool execution. At $1.50 per million tokens it sits in the middle of the field: pricier than the bargain multimodal tier, cheaper than the flagships.
For Meta, open models are strategy, not charity: the Llama family taught the industry that open weights earn influence, and Muse Glimmer extends that playbook into the agentic era under Meta's own community licence. If your interest is running a capable agent stack on hardware you own, this is one of the few models here designed explicitly with that in mind.
What can you do with Muse Glimmer 30B?
A distillation of a much larger model, aimed at agents on modest hardware:
- Self-hosted agent stacks — 30B is a size a serious workstation can actually serve.
- Tool-calling workflows: browsing, shell work, API orchestration with steady multi-turn behaviour.
- Running agent workflows on modest hardware, as the model's announcement describes.
- Studying distillation — an open window into how a big model's habits survive shrinking.
- Long-horizon tasks inside its roughly 100,000-word context window.
- Private or offline deployments, where the zero-data-retention terms matter.
How does Muse Glimmer 30B compare?
Its nearest rival is OpenAI's GPT-OSS 120B — the other compact open agent — which is cheaper per million tokens ($0.60 versus $1.50) and reasons step by step, though with a slightly smaller context. The choice between them is mostly temperament and licence: Apache-2.0 simplicity and OpenAI tooling versus Meta's distillation pedigree and compact footprint.
Against the careful mid-size models — Google's Gemma 4 31B, Alibaba's Qwen3.8 27B — Muse Glimmer is the agent-first one: they write and reason, it acts. And against Meta's own scale it is the people's version: whatever the parent model could do, this is the part you can run at home.
Is Muse Glimmer 30B right for you?
Muse Glimmer 30B is the model to pick when the hardware is part of the decision. If you are self-hosting, running agents on workstations rather than data centres, or building something that has to run on machines you already own, a distilled 30B with native tool support is exactly the shape you want, and the Meta Superintelligence Labs pedigree shows in the training. The zero-data-retention terms are a genuine plus for private deployments, and the community licence keeps the weights available to tinker with.
For chat-only users the calculus is different. At $1.50 per million tokens it costs more than several stronger-feeling general models, and its 131k context is modest by this catalogue's standards — if you are not exploiting the small footprint, you may be paying for a feature you do not use. GPT-OSS 120B is the natural alternative for agent work at lower cost, and the Gemma pair covers careful and quick work respectively.
Where it earns devotion is the middle: people who want a capable agent they can actually own, from a lab that publishes what it learns. That is a narrower brief than most models here claim, and Muse Glimmer fills it without pretending to be anything else.
Frequently asked questions
Who makes Muse Glimmer 30B?
Meta Superintelligence Labs — the company's frontier AI research effort. Meta has a long history of releasing open models, from the Llama family onward, and Muse Glimmer continues that practice into the agentic era. The Superintelligence Labs branding signals where it sits — the company’s frontier effort, shipping open weights as strategy rather than as overflow.
What does it mean that Muse Glimmer is distilled?
Distillation means the model was shrunk down from a much larger Meta system: the big model's behaviour was compressed into 30 billion parameters so it can run on ordinary hardware. You keep much of the parent's ability at a fraction of the size and cost. Distillation is why it keeps the parent’s judgement on small decisions while giving up raw depth — a trade that suits agent loops, which need steadiness more than brilliance.
What is the context window of Muse Glimmer 30B?
131k tokens — roughly 100,000 words. That is a window sized for realistic agent loops and long documents rather than the million-token archives the biggest models carry.
Does Muse Glimmer 30B support image input?
No. In the UsingOpen catalog and comparison data, Muse Glimmer 30B is configured as a text-only agent model with native tool calling. Tool support and the 131k window are the core design points for local agentic execution; it is served here as a text-and-tool agent model.
Is Muse Glimmer 30B good for agent work?
Yes — that is its stated purpose: agents on modest hardware. It supports tool calling natively, keeps a sensible context window for long loops, and is small enough to self-host if you would rather run it yourself. Meta’s own framing is agents on modest hardware, and the 30B size is chosen for exactly that: small enough to serve locally, large enough to trust with a loop.
Is Muse Glimmer 30B free to chat with on UsingOpen?
Yes. Pick it in the model picker and start chatting — no account, no payment. History is saved only on your device, and the picker shows live availability. Trying an agent-focused model without setting up any hardware is the cheapest way to find out whether the agentic temperament suits your work.
Related models on UsingOpen
Models most often compared with Muse Glimmer 30B on UsingOpen:
- GPT-OSS 120B by OpenAI: the open-weight OpenAI reasoning model — 117B total, 5B awake, built to reason step by step.
- Gemma 4 31B by Google: the dense 31B Gemma — strong writing, picture input and visible thinking.
- Qwen3.8 27B by Alibaba: the newest mid-size Qwen — a dense 27B that thinks before it answers.
Head-to-head comparisons
See how Muse Glimmer 30B lines up with the models people most often weigh it against:
- Muse Glimmer 30B vs Gemma 4 31B — Meta's compact agent model against Google's cheaper writer and image reader.
Model availability changes as upstream capacity comes and goes — the picker on the chat page shows what is ready right now.
Chat with Muse Glimmer 30B free on UsingOpen — no account, no sign-up
← All open-weight models · Compare models · Free AI chat · FAQ