Meta Muse Glimmer 30B: How to Run It Locally and Is It Worth It?
Meta's open-weight Muse Glimmer 30B rivals Qwen 3.6 27B on agent benchmarks. Here's the hardware, quantization, and setup to run it yourself.
Drafted with Claude from source material, checked by automated verification, and reviewed before release. How we make these.
What is Muse Glimmer 30B?
Muse Glimmer is a 30 billion parameter open-weight language model released by Meta under an Apache 2.0 license. It’s a dense model, not a mixture-of-experts design, which means all 30B parameters activate on every forward pass rather than routing through a smaller subset. Meta built it specifically for agentic work: multi-step reasoning, tool use, long task trajectories, and compatibility with coding harnesses and agent frameworks. It’s positioned as a direct competitor to Qwen 3.6 27B, and Meta’s own benchmark charts pit the two head to head, along with comparisons against Gemma 4.
TL;DR
- Meta released Muse Glimmer, a 30B dense open-weight model under Apache 2.0, marking its return to shipping open weights after the Llama era wound down.
- The model is built for agents, with training and benchmarks emphasizing tool use, multi-step reasoning, and long-horizon task trajectories rather than pure chat quality.
- Meta shipped an official 4-bit quantized version sized to fit on 24GB to 32GB cards, targeting RTX 3090/4090/5090-class GPUs and cards like the AMD 9700.
- The quantized build supports speculative decoding (Meta calls it D-Flash), which speeds up generation by having a smaller draft model propose tokens the main model verifies.
- Meta demoed the model running on a MacBook Pro with 64GB of unified memory, something that would have been unthinkable with the original Llama releases.
- Benchmarks show Glimmer beating Qwen 3.6 27B on many tasks but not all, with Meta racing to release before Qwen’s 3.8 27B update lands.
- Meta also confirmed it will open-weight the larger Muse Spark 1.2 model, which independent tracking already places near Claude Opus 4.5 in capability.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How was Muse Glimmer trained?
Meta’s accompanying report describes a training pipeline that leans heavily on distillation rather than training from scratch on raw web data. Glimmer’s pre-training draws on outputs from the larger Muse Spark model, essentially using the bigger model as a teacher. Post-training combines on-policy distillation with reinforcement learning, a recipe that mirrors what other labs (including Thinking Machines’ recent small-model work) have been experimenting with. The difference here is scale and intent: Glimmer is small enough to run on consumer hardware while still inheriting behavior from a much larger, more capable sibling.
This approach matters for anyone evaluating the model. A distilled model built on top of a stronger teacher can punch above its parameter count on the specific behaviors the teacher was good at, but it may also inherit blind spots or stylistic quirks from that teacher rather than developing them independently from raw data.
What hardware do you need to run it locally?
Meta shipped an official 4-bit quantized version of Muse Glimmer sized deliberately to fit within 24GB to 32GB of VRAM. That puts it squarely in range of:
- Nvidia RTX 3090 (24GB)
- RTX 4090 (24GB)
- RTX 5090 (32GB)
- AMD Radeon AI PRO R9700 / similar 32GB workstation cards
Meta didn’t just cram the model into that memory ceiling. The quantized build leaves headroom for a reasonably sized KV cache, the memory structure that stores attention context as you generate longer outputs. That headroom matters in practice: a model that barely fits at 4-bit with no cache room becomes nearly unusable once you start feeding it long agent transcripts or multi-turn conversations.
On the Mac side, Meta demonstrated the model running on a MacBook Pro with 64GB of unified memory. Apple Silicon’s unified memory architecture lets the GPU and CPU share the same pool, so a 64GB machine has plenty of room for a 30B model at 4-bit quantization plus cache. Token generation speed on Mac will trail a dedicated 5090, but the fact that a laptop can run this at all is a meaningful shift from the Llama era, when running large models locally usually meant a multi-GPU workstation.
What is speculative decoding and why does it matter here?
Muse Glimmer’s quantized release includes speculative decoding support, which Meta refers to as D-Flash. Speculative decoding works by pairing the main model with a smaller, faster draft model. The draft model guesses several tokens ahead, and the main model verifies them in a single batched pass instead of generating token by token. When the guesses are right, which happens often for predictable text, you get a substantial speedup with no loss in output quality, since the main model still does the final verification.
Building this in at release time signals that Meta expects Glimmer to be used in latency-sensitive agent loops, where a task might involve dozens of sequential model calls (planning, tool calls, verification, replanning). Shaving latency off each step compounds quickly over a long agent trajectory.
How does Muse Glimmer compare to Qwen 3.6 27B?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Meta’s own benchmark suite frames Glimmer as a Qwen 3.6 27B competitor rather than a categorical leap forward. Glimmer wins on a number of benchmarks and loses on others, roughly the profile you’d expect from two models trained with different data mixes and post-training recipes at similar scale. It also outperforms Gemma 4 by a wider margin across most of the tests Meta published.
The comparison is complicated by timing. Qwen’s next iteration, Qwen 3.8 27B, is expected within days of Glimmer’s release. That raises the obvious question of whether Meta rushed Glimmer out specifically to claim a benchmark win against 3.6 before the goalposts moved. Even if that’s the case, the benchmarks themselves are real and the model is genuinely usable today, so the comparison is still useful for anyone deciding what to run right now rather than waiting for the next release cycle.
Is Muse Glimmer worth running instead of Qwen?
For most builders, the honest answer is: try both. Glimmer is Meta’s first standalone open-weight release under its rebuilt research team, and first releases in a new model family typically improve fast as the team iterates on post-training and distillation techniques. Qwen’s models have had several release cycles to mature. If your workload is agent-heavy (tool calling, long trajectories, harness compatibility), Glimmer’s training was explicitly optimized for that, which may show up as more reliable behavior in agent frameworks even where raw benchmark numbers are close.
The Apache 2.0 license is also worth noting on its own terms. It’s a permissive license with no usage restrictions on commercial deployment, which removes a common source of friction when choosing a model for production agent pipelines.
What comes next from Meta’s open model push?
Mark Zuckerberg confirmed alongside the Glimmer announcement that Meta will release open weights for the larger Muse Spark model (version 1.2) as well. That model has already been tracked by third-party benchmarking outfits like Artificial Analysis, which places its Intelligence Index score near Claude Opus 4.5, a notably strong result for a model that may soon be downloadable and runnable outside a hosted API. Meta has also recently shipped Muse Code, a coding-focused variant of Muse Spark, though it’s not yet confirmed whether that one will be open-weighted.
Taken together, these releases suggest Meta’s newer research team (which reportedly includes former Gemini reasoning leadership) is shipping on a real cadence rather than a one-off. For anyone building locally-run agent tooling, that’s the more durable story than any single benchmark table: a major lab is back in the open-weights business with models sized for the hardware people actually own.
Frequently Asked Questions
What license is Muse Glimmer 30B released under?
Apache 2.0, a permissive open-source license that allows commercial use, modification, and redistribution without the usage restrictions found in some other “open” model licenses.
How much VRAM do I need to run Muse Glimmer locally?
Meta’s official 4-bit quantized version is sized to fit on 24GB to 32GB of VRAM, matching cards like the RTX 3090, RTX 4090, RTX 5090, and AMD’s 32GB workstation cards, with headroom left for the KV cache.
Can I run Muse Glimmer on a Mac?
Yes. Meta demonstrated the model running on a MacBook Pro with 64GB of unified memory, though generation speed will be slower than on a dedicated Nvidia GPU like the 5090.
Is Muse Glimmer better than Qwen 3.6 27B?
One coffee. One working app.
You bring the idea. Remy manages the project.
It beats Qwen 3.6 27B on a number of published benchmarks but not all of them. It’s best understood as a close competitor at similar scale rather than a clear outright winner, and Qwen’s 3.8 27B update is expected shortly after Glimmer’s release.
Is Muse Glimmer a mixture-of-experts model?
No. It’s a dense 30B parameter model, meaning all parameters are active on every inference pass, unlike mixture-of-experts architectures that route through a subset of parameters per token.