Meta Superintelligence Labs dropped Muse Glimmer today — a 30-billion-parameter agentic model under Apache 2.0 that fits on a single 24GB consumer GPU. After years of “local agents are coming,” the gap between what’s theoretically possible and what you can run on hardware you own just collapsed by one more step.
What You Can Actually Run It On #
The marketing line is “runs on one consumer GPU.” The real answer depends on which quantized variant you pick:
K-Quant-17GB(~17GB file): Requires 24GB of VRAM or Mac unified memory — an RTX 5090, RTX 4090 (tight), M4-Max, or M5-Max. Accuracy loss: 1.0%.K-Quant-Dynamic(~20GB file): Requires 32GB. Accuracy loss: 0.2%.** BF16 reference**: 1x H100 80GB. For evaluation and fine-tuning only.
The practical floor is a 24GB GPU. If you’re on a 16GB card, wait for community quants in the 12–14GB range — or run it hosted on Together AI or OpenRouter while you wait.
Getting started is one command with llama.cpp:
curl -LsSf https://llama.app/install.sh | sh
llama serve -hf meta-models/Muse-Glimmer-30B-GGUF
That exposes an OpenAI-compatible endpoint at http://localhost:8080/v1
. Point any tool that accepts an OpenAI API URL at it and you’re running a local agentic model. For a 3.1x decode speedup on RTX 5090 hardware, add speculative decoding:
llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --spec-type draft-dflash --spec-draft-n-max 15
Ollama 0.32.7 shipped the same day with Muse Glimmer support built in. LM Studio and vLLM (via the Transformers backend) also work on day one.
Where It Wins — and Where It Doesn’t #
Meta benchmarked Glimmer against Gemma4-31B and Qwen3.6-27B. Here is the condensed version:
| Benchmark | Muse Glimmer | Qwen3.6-27B | Gemma4-31B |
|---|---|---|---|
| MCP Atlas (agentic) | 75.5 | 62.5 | 54.2 |
| SWE-Bench Pro (coding) | 51.2 | — | 36.9 |
| AIME 2026 (reasoning) | 94.7 | 94.1 | 89.2 |
| TerminalBench 2.1 | 51.7 | 60.7 | — |
| OSWorld-Verified | 65.9 | 75.6 | — |
The pattern is clear: Glimmer is the pick for MCP-based orchestration, multi-step coding agents, and reasoning tasks. Qwen3.6 still leads on sustained terminal work and computer control. No model sweeps every row — and Meta’s benchmarks use different sampling settings across comparisons, so reproduce any result that matters to your stack before committing hardware.
The Qwen3.8 Timing Problem #
Hacker News called it within hours: Meta dropped Glimmer now because Qwen 3.8’s open weights are expected this week. Alibaba released Qwen3.8-Max via API on August 3. The 27B open-weight version is expected the week of August 10. Community consensus is that Qwen3.8 will close most of Glimmer’s leads and may flip several rows in the table above.
This is not a reason to ignore Glimmer — it’s a reason to run your own evals before finalizing any architecture decision. Download Glimmer today, test it on your actual workloads, and wait for Qwen3.8 open weights before picking a winner. Both will be Apache 2.0 and both will run on the same hardware.
Why Local Inference Actually Matters Here #
The conversation about local LLMs is often framed as hobbyist preference over cloud convenience. That framing is wrong. Enterprise on-premises AI inference has risen from 12% in 2023 to 55% in 2026. For teams in healthcare, finance, or government, data residency is frequently a legal requirement. Under GDPR Article 28, local inference eliminates the data processor obligation for the model provider entirely.
Muse Glimmer is the first model in this size class — 30B, genuinely agentic, multimodal — that fits on hardware a team can physically own. Apache 2.0 means no usage fees and no terms-of-service surprises. That combination has real commercial value beyond benchmark bragging rights.
Before You Put It in Production #
Prompt injection attacks succeeded 28.4% of the time in the Siren AgentDojo evaluation. That is not a reason to avoid Glimmer; it is a reminder that no local model substitutes for tool-call allowlists, network isolation, and human-in-the-loop confirmation on high-stakes actions. Glimmer also does not process audio from video inputs and does not generate images or audio. For most agentic coding and document workflows those are not blockers — but know before you design around them.
Get Started #
Weights are at meta-models/Muse-Glimmer-30B on Hugging Face. GGUF quantizations at meta-models/Muse-Glimmer-30B-GGUF
. Full technical architecture in the Hugging Face model blog. The official announcement is on Meta’s research blog.
Run it, test it against your actual workloads, and revisit the decision when Qwen3.8 open weights land.