# Meta Superintelligence Labs Returns to Open Weights With a 30B Agent Model Under Apache 2.0

Muse Glimmer is dense rather than mixture-of-experts, carries a context window above 128,000 tokens, and fits under 20 gigabytes at four-bit quantization.

- Published: 2026-08-11T06:26:24.055Z
- Canonical: https://polylog.news/ai/2026-08-11/meta-superintelligence-labs-returns-to-open-weights-with-a-3
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Ollama](https://ollama.com/blog/muse-glimmer), [Polylog editors](https://polylog.news), [VentureBeat](https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now), [AMD](https://www.amd.com/en/blogs/2026/run-meta-muse-glimmer-30b-on-amd-ryzen-ai-max-and-radeon-gpus.html)

Meta Superintelligence Labs released [Muse Glimmer](https://ollama.com/blog/muse-glimmer), a 30-billion-parameter dense multimodal model under an Apache 2.0 license, and it is the group's first open-weights release. The stated target is narrow and practical: local coding agents that run continuously on a developer's own hardware rather than a hosted frontier model billed per token.

The engineering choices follow from that target. Full-precision weights need more than 55 gigabytes, but a four-bit quantization brings the model under 20 gigabytes and lets a complete setup fit in 24 or 32 gigabytes of memory, which is one consumer graphics card or a well-specified laptop. Meta paired the release with DFlash speculative decoding (a technique that predicts several tokens ahead to speed up generation), reporting a 3.1 times generation speedup on an Nvidia RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max. Support was available at launch across Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang, and [AMD published its own guide](https://www.amd.com/en/blogs/2026/run-meta-muse-glimmer-30b-on-amd-ryzen-ai-max-and-radeon-gpus.html) for Ryzen AI Max and Radeon parts.

The license is the substantive change. Meta's previous open releases carried custom terms with usage caps, and [VentureBeat noted](https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now) that Apache 2.0 permits commercial use, modification and redistribution without a separate grant. A Russian-language technical channel [described the release](https://t.me/ai_machinelearning_big_data/10679) in the same terms, emphasizing continuous local operation of agent systems as the design point.

What has not been established is how Muse Glimmer compares to the Qwen and GLM open-weight models that currently occupy the same size class. Meta has published speedup figures and hardware footprints, both of which are easy to verify, but comparative agentic and coding scores against those competitors are not yet independently reproduced. A 30-billion-parameter dense model is also an unusual choice in a period when most releases at this capability level use sparse mixture-of-experts routing, and dense inference costs more per token at scale even as it simplifies local deployment.

## What this means

Meta is competing for the local-agent developer, not the hosted-API customer, and Apache 2.0 removes the legal review that previously kept its models out of commercial products. That is a direct threat to the low end of the paid-API market, because a coding agent that runs unmetered on a 24-gigabyte card has a marginal cost of electricity rather than tokens. Alibaba's Qwen line and Zhipu's GLM line lose the license advantage they held over Meta, while Ollama, LM Studio and the consumer graphics vendors gain from any shift of agent inference off rented cloud capacity.

## What to watch

- Independent coding and tool-use benchmark results comparing Muse Glimmer against Qwen and GLM models of similar size, which will show whether the license change is backed by competitive capability.
- Whether Meta releases larger Muse models under the same license, which would indicate this is a strategy rather than a single low-risk trial.
- Adoption of DFlash speculative decoding by other open models, since a portable three times speedup on consumer hardware would compress local inference costs broadly.