Meta Went Back to Open Weights
The licence is the news. Apache 2.0 means what it says, which Meta's previous releases did not.
Founder, Automation Squad ·
The short answer
Meta Superintelligence Labs released Muse Glimmer on August 10, 2026 — a 30B dense multimodal agentic model under the Apache 2.0 licence, rather than Meta's previous community licence. At 4-bit quantisation it fits under 20GB, running in a 24GB or 32GB envelope alongside KV cache and supporting components.
Running Muse Glimmer on hardware you already have
Sizing, the runtimes it ships on, and what the licence actually permits. Aimed at getting one working generation, not at a production deployment.
Check your VRAM against the real envelope, not the file size
Under 20GB at 4-bit is the model. You also need room for KV cache, the perception encoder and the speculative-decoding drafter, which is why the practical figure is a 24GB or 32GB card.
Start with Ollama or LM Studio if you have never run a local model
Both are supported at launch and both remove the setup problem entirely. vLLM and SGLang are the right answer for serving; they are the wrong answer for finding out whether this is useful to you.
Test it on the job you currently pay per token for
The economics of local inference only make sense at volume. Take your highest-frequency, lowest-sensitivity task and run it both ways on the same fifty inputs.
Note that Apache 2.0 genuinely is permissive
This is the meaningful change. Meta's earlier open releases carried a community licence with conditions that mattered commercially. Apache 2.0 does not carry those, which makes this usable in places the previous releases were not.
Compare it against the other 30B that shipped the next day
NVIDIA released a 30B Mixture-of-Experts model on August 11 under a different licence. Two comparable options a day apart is unusual and worth an hour of testing rather than picking on brand.
| Detail | Value |
|---|---|
| Parameters | 30B dense, multimodal, agentic |
| Licence | Apache 2.0 — genuinely permissive |
| Size at 4-bit | Under 20GB |
| Practical envelope | 24GB or 32GB with KV cache and encoder |
| Runtimes at launch | vLLM, SGLang, Ollama, LM Studio |
| Coming shortly after | llama.cpp, MLX, ExecuTorch |
| Benchmarked against | Gemma4-31B, Qwen3.6-27B |
| Weights | meta-models/Muse-Glimmer-30B (plus GGUF) |
Take it with you
MUSE GLIMMER — running it locally
Meta, Aug 10 2026, Apache 2.0
Source: https://automationsquad.com/news/meta-muse-glimmer-open-weights/
THE MODEL
30B dense, multimodal, agentic
Licence ................ Apache 2.0 (genuinely permissive)
Size at 4-bit .......... under 20GB
Practical envelope ..... 24GB or 32GB card
(model + KV cache + perception encoder
+ DFlash speculative-decoding drafter)
Weights ................ meta-models/Muse-Glimmer-30B (+ GGUF)
RUNTIMES
At launch .............. vLLM · SGLang · Ollama · LM Studio
Shortly after .......... llama.cpp · MLX · ExecuTorch
GETTING ONE GENERATION OUT
[ ] Check VRAM against the ENVELOPE, not the file size
[ ] New to local models? Start with Ollama or LM Studio.
(vLLM/SGLang are for serving, not for finding out if it's useful)
[ ] Pick my highest-frequency, lowest-sensitivity paid task
[ ] Run the same 50 inputs locally and hosted. Count disagreements.
[ ] Compare against NVIDIA's 30B MoE from Aug 11 — different licence,
different architecture, shipped a day apartMeta Superintelligence Labs released Muse Glimmer on August 10, 2026 — a 30B dense multimodal agentic model — under the Apache 2.0 licence. Mark Zuckerberg published a long essay on open source and personal intelligence the same day.
The facts: at 4-bit quantisation the model fits under 20GB, running in a 24GB or 32GB envelope once KV cache, the perception encoder and a DFlash speculative-decoding drafter are accounted for. It was benchmarked against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning evaluations. It ships on vLLM, SGLang, Ollama and LM Studio, with llama.cpp, MLX and ExecuTorch support following shortly. Weights are published as meta-models/Muse-Glimmer-30B, with GGUF builds alongside.
Automation Squad's take: the licence is the story, and it is being under-reported in favour of the benchmarks. Meta's previous open releases carried a community licence with conditions that genuinely constrained commercial use, which meant "open" required a lawyer to interpret. Apache 2.0 does not. That single change moves this from a model you can experiment with to a model you can ship on. The sizing is the second thing worth noting — 30B at 4-bit landing under 20GB is a deliberate choice to fit a card that individuals and small teams already own, which is the same bet NVIDIA made twenty-four hours later with a different architecture.
Run this now: check your GPU's memory against the 24GB-or-32GB working envelope rather than the model file size, because the difference is where people get caught. If you have never run a model locally, install LM Studio or Ollama and get one generation out before optimising anything. Then take the single task you currently pay per token for most often — high frequency, nothing confidential — and run fifty real inputs through both the local model and your hosted one. The disagreement count is your answer, and it is worth more than any published benchmark.
Questions people are asking
- What licence is Muse Glimmer released under?
- Apache 2.0 — a genuinely permissive licence, rather than the community licence Meta applied to earlier open releases. That change is the most commercially significant part of the announcement.
- What hardware do I need to run Muse Glimmer?
- At 4-bit quantisation the model fits under 20GB, but the practical requirement is a 24GB or 32GB card once you account for KV cache, the perception encoder and the speculative-decoding drafter.
- How do I run it without setting up infrastructure?
- It shipped with support for Ollama and LM Studio alongside vLLM and SGLang. The first two require essentially no setup, which makes them the right starting point for evaluating whether it is useful to you.
Sources
Further reading
- Meta — Introducing Muse Glimmer — Model card, benchmarks, and the runtime support matrix.
Last checked August 10, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].