---
title: Meta's New 30B Model Runs on Consumer Hardware Under the Apache 2.0 License
description: Meta's new 30B open model, Muse Glimmer, runs local agents on a 24GB card and ships under Apache 2.0 instead of the Llama Community License.
author: Darie Nani (Editor-in-Chief)
date: 2026-08-10T19:28:32.247Z
updated: 2026-08-10T19:28:50.354Z
canonical: https://www.sovereignmagazine.com/article/meta-muse-glimmer-open-30b-apache-2
image: https://cdn.nanimediahouse.com/unsplash-logo-WR6qHgdWS-Y.jpg
categories: Artificial Intelligence
content_type: News
region: Global
publication: Sovereign Magazine
schema_type: Article
---

Meta has released a 30 billion parameter model called Muse Glimmer under the [Apache 2.0 license](https://huggingface.co/meta-models/Muse-Glimmer-30B), not the Llama Community License that covers its Llama models, and it is designed to run agents on a machine a developer already owns. The repository on Hugging Face is not gated, so there are no terms to accept before downloading it.

Muse Glimmer is built for agent work rather than conversation. Meta Superintelligence Labs distilled it from Muse Spark, [the model Meta has promoted as its most capable coding system](https://www.sovereignmagazine.com/article/ai-models-three-labs-breached-companies-testing), and gave it a perception encoder so it can read screenshots, charts and documents as part of a task. Meta's model card says it runs locally without cloud infrastructure or network access.

## Llama's User-Count Clause and Download Gate Are Both Gone

Meta wrote the Llama Community License itself, and it is not a standard open-source license. [Clause 2 of the Llama 3.3 agreement](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct), headed Additional Commercial Terms, says a licensee whose products or services pass 700 million monthly active users in the preceding calendar month must request a license from Meta, which Meta may grant in its sole discretion. Llama repositories on Hugging Face are gated, so a developer accepts those terms before the download begins.

Muse Glimmer carries neither condition. Few developers were ever going to cross that threshold, but the gate and the discretionary clause applied to all of them.

NVIDIA's developer blog describes the release as Meta returning to the open source ecosystem. The release comes five days after the [White House said it will not put open-weight models, Meta's Llama among them, through its new security testing program](https://www.sovereignmagazine.com/article/white-house-open-weight-ai-models-security-testing).

## The Compressed Model Fits Under 20 Gigabytes

Full-precision BF16 weights run to about 58GB, beyond what a desktop holds. Meta's own quantized build compresses the weights to roughly 4-bit, which brings the language model under 20GB and leaves headroom for the KV cache, the perception encoder and the speculative decoding drafter to run at the same time within a 24GB or 32GB envelope.

Meta measured what the compression costs. Averaged across 15 common benchmarks, accuracy falls 0.2 percent on the build aimed at 32GB and 1.0 percent on the one aimed at 24GB. Unsloth, which worked with Meta and Hugging Face on the llama.cpp implementation, puts the practical entry point at 18GB of combined RAM and VRAM, on Mac as well as GPU and CPU systems, and lists the 4-bit build at about 17GB.

## A Drafter Model Roughly Triples Throughput on an RTX 5090

Muse Glimmer ships with a small companion model, based on a technique called DFlash, that proposes blocks of 16 tokens at once for the main model to verify in parallel instead of generating one token at a time. With that drafter and the 24GB build, Meta measured 233.4 tokens per second on an Nvidia RTX 5090, up from 74.9 without it, and 37.8 on an Apple M4 Max, up from 23.7.

AMD ran its own early tests on Windows through llama.cpp with the Vulkan backend and reported up to 24 tokens per second on a Ryzen AI Max+ 395 processor and up to 53 on a single Radeon AI PRO R9700 card, figures it calls preliminary. AMD's case for keeping the work local is the latency and the recurring token bills of a cloud-only setup.

## The Model Retries When a Tool Call Fails

When a tool call fails or returns something unexpected, Meta's model card says, the model diagnoses the error and retries rather than stopping. It works inside agent orchestration setups including OpenClaw and Hermes Agent. Unsloth notes that the ability to resume work across a long session comes from the agent harness around the model, not from the model itself.

NVIDIA puts the context window above 120,000 tokens and lists support across the GeForce RTX 5090, DGX Spark, DGX Station and Jetson, deployable through NIM containers, SGLang and vLLM. Every parameter activates for each token, which NVIDIA says gives steadier latency and better long-context coherence than a mixture-of-experts design.

On Meta's own card the model scores 76 on SWE-bench Verified and 83.5 on GPQA Diamond. It was trained on data from more than 100 languages, and its knowledge cutoff is 4 January 2026.

## FAQ

**Q: Is the Apache 2.0 license safe for commercial use?**
Apache 2.0 carries no user-count trigger and no requirement to ask Meta for anything, which is the practical difference from the Llama Community License. That license requires a company above 700 million monthly active users to request permission from Meta directly, granted at Meta's discretion.

**Q: How much memory do you need to run Muse Glimmer locally?**
Unsloth lists 18GB of combined RAM and VRAM as the entry point, on Mac as well as GPU and CPU systems, with the 4-bit build at about 17GB. Meta's own quantized build targets a 24GB or 32GB envelope for the language model, the perception encoder and the speculative decoding drafter running together. Full-precision weights need about 58GB.

**Q: What do you give up by running the 4-bit version?**
Meta puts the cost at 1.0 percent of accuracy on the build aimed at 24GB and 0.2 percent on the one aimed at 32GB, averaged across 15 common benchmarks.