Overview

Muse Glimmer is a 30-billion-parameter open-weight multimodal model Meta released on August 10, 2026, built around one job: running agents on hardware you already own. Meta's framing is tool use and failure recovery rather than conversation — the model is meant to call things, notice when a call comes back wrong, and keep going, instead of producing a tidy paragraph about what it would have done. Weights are on Hugging Face under Apache 2.0, so commercial use and redistribution don't require a vendor-specific community license. Zuckerberg announced it on X the same day it landed, and Bloomberg covered it.

The audience is anyone who wants an agent loop that doesn't meter per token or send working context to someone else's endpoint: local-first developers, teams under data-residency or client-confidentiality constraints, and people wiring agents into desktop or internal workflows where a per-call bill scales badly. Size is the design decision worth noticing. Thirty billion parameters at 4-bit quantization lands in the rough neighbourhood of 16–20GB of memory, which puts it on a single high-end consumer GPU or a well-specced Apple-silicon machine — clearly above the 7–8B models people run as toys, and clearly below anything that needs a rack. You still supply the runtime yourself: Ollama, LM Studio, vLLM or similar.

The context makes this more interesting than the spec sheet. Meta's previous notable release, Muse Spark, went the other direction entirely — a hosted, metered Model API with a cheaper tier priced in training rights on your code. Glimmer is Meta putting agentic capability back on disk for free. My read: at two days old there are no independent evaluations, so "failure recovery" is a training objective and a launch claim, not a reproduced result, and nobody has yet reported how it holds up over long multi-step runs where small models usually come apart. Download it if you already run local models and want a better agent backbone than an 8B chat tune. Don't retire a frontier hosted model on the strength of an announcement post.

Is Muse Glimmer free?

Yes. Muse Glimmer is free to use.

What the free tier covers: The entire model is free. Apache 2.0 open weights downloadable from Hugging Face, with no account tier, rate limit or per-token metering — inference cost is whatever your own hardware or cloud GPU rental costs.

Listed pricing: Free — Apache 2.0 open weights on Hugging Face.

Pricing on this page has not been re-verified. The entry was last edited on 2026-08-11, and no separate pricing check has been run since. Treat the figures as a record of what was published then and confirm on the official site.

Muse Glimmer pricing

Plan	Price	What's included
Open weights	Free	Full model weights on Hugging Face under Apache 2.0 — commercial use, modification, fine-tuning and redistribution permitted. Your only cost is the hardware you run it on. Meta has not announced a hosted first-party endpoint for Glimmer; check the model card for current license terms.

Open weights Free

Full model weights on Hugging Face under Apache 2.0 — commercial use, modification, fine-tuning and redistribution permitted. Your only cost is the hardware you run it on. Meta has not announced a hosted first-party endpoint for Glimmer; check the model card for current license terms.

Pricing on this page has not been re-verified. The entry was last edited on 2026-08-11, and no separate pricing check has been run since. Treat the figures as a record of what was published then and confirm on the official site.

Is Muse Glimmer worth it?

You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.

The 8.4/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with Muse Glimmer. How we verify.

Worth it if

The strengths recorded against this entry.

Apache 2.0 rather than a restrictive community license — commercial use, fine-tuning and redistribution are all permitted
Runs on hardware you own, so agent context and files never leave the machine and a chatty tool loop costs nothing per call
30B is the useful middle: meaningfully more capable than 7–8B local models, still small enough for one high-end consumer GPU at 4-bit
Explicitly trained for tool use and failure recovery rather than being a chat model with function-calling bolted on
Immediately usable through the existing Hugging Face ecosystem — Ollama, LM Studio and vLLM users need no new tooling

Not worth it if

Any one of these blocks your use case.

No independent evaluations yet — everything about its agent performance rests on Meta's own launch claims as of August 11, 2026
Still too heavy for a typical 8GB laptop GPU; you need a high-end consumer card or a well-specced Apple-silicon machine
It's a model, not a product — you supply the runtime, the harness, the quantization choice and all the ops
Meta has announced no hosted first-party endpoint, so there's no zero-setup path to try it before committing hardware

Key features

30B open weights

Apache 2.0 weights published on Hugging Face, so the model can be downloaded, fine-tuned, redistributed and run commercially without a vendor-specific community license or usage threshold. At 4-bit quantization a 30B model sits roughly in the 16–20GB range, which is single-consumer-GPU territory rather than datacentre territory.

Agentic tool use

Meta positions the model for calling tools and chaining multi-step tasks rather than for chat, including recovering when a tool call fails instead of stalling or hallucinating a result. This is the part that separates it from general-purpose small models that technically emit function calls but fall apart mid-loop.

Local-first deployment

Running on your own machine means prompts, files and intermediate agent state never leave it, and there is no per-token bill on a loop that might fire hundreds of calls. The tradeoff is that you own the runtime, the quantization choice and the ops.

Multimodal input

The model accepts more than text, so screenshots and images are legitimate agent input rather than something you route to a second service. Meta has not published a detailed breakdown of which modalities are supported at what quality, so verify against the model card before designing around it.

How it compares

Tool	Best for	Pricing	Score
Muse Glimmer	—	Free — Apache 2.0 open weights on Hugging Face	8.4/10
n8n	Developer-adjacent ops and engineering teams who want AI agents and internal integrations running on infrastructure they control, with server costs instead of a per-task meter.	Free self-hosted + Cloud from ~$20/mo	8.7/10
Make	Teams building multi-step, branching automations who want to see the whole flow on a canvas rather than describe it as a list of steps.	Free tier with 1,000 credits/mo; paid from $9/mo for 10,000 credits	8.7/10
Fathom	Founders, consultants and small sales teams living in Zoom or Google Meet who want unlimited call recording and CRM-synced summaries without a visible bot joining client calls.	Freemium — generous free tier; paid Premium and Team plans priced per seat (check website for current rates)	8.5/10