META MUSE GLIMMER — KEY FACTS (RELEASED AUGUST 10, 2026)
● Download: Hugging Face — meta-ai/Muse-Glimmer (Apache 2.0, free)
● Parameters: 30B dense (28B text decoder + 1.8B ViT vision encoder)
● Hardware: 24GB VRAM GPU (Q4_K_M GGUF quant) or Apple Silicon Mac
● License: Apache 2.0 — commercial use, fine-tuning, redistribution permitted
● Inference: Ollama · llama.cpp · LM Studio · vLLM · SGLang · Unsloth · ExecuTorch
● Speed: 3.1× speedup via DFlash speculative decoding drafter
● Context: 131,072 tokens
● Languages: 100+
● Vision: Text + image input, text output. No audio, video processed as frames.
● Knowledge cutoff: January 4, 2026
● MCP Atlas: 75.5 (vs Gemma4-31B 54.2, Qwen3.6-27B 62.5) — #1
● SWE-Bench Pro: 51.2 — strong agentic coding
● AIME 2026: 94.7 — strong reasoning
● Security note: 28.4% Siren AgentDojo attack success rate — always run in container/sandbox
Why This Release Is Different
Per Meta's official model page, Muse Glimmer is built for "always-on local agents" — the specific workload where running locally changes the economics entirely. As DEV Community's analysis explains, the current API-based agent economics mean "every tool call, every retry, every failed attempt is tokens on someone else's infrastructure, billed per request. A local agent inverts that. The cost becomes hardware you already own, the data never leaves the device, and the agent keeps working in a tunnel, on a plane, or on a client site with no network." At 30B parameters under Apache 2.0 running on a 24GB GPU, Muse Glimmer is the first model to make that inversion practical for most development teams — without requiring the 8× H100 minimum of Kimi K3 or the MIT license restriction of DeepSeek V4 Flash.
As VentureBeat's coverage notes, Mark Zuckerberg also announced that Muse Spark 1.2 weights are coming soon — which would put a US frontier model into open circulation for the first time. Per AI Weekly's analysis, "if a 30B Apache-licensed agent model that actually runs on a 24GB card holds up to Meta's benchmarks, the economics of paid agent APIs narrow fast."
Benchmark Results — What It Wins and What It Loses
| Benchmark | Muse Glimmer | Gemma4-31B | Qwen3.6-27B |
|---|
| MCP Atlas (agentic) | 75.5 🥇 | 54.2 | 62.5 |
| SWE-Bench Pro | 51.2 🥇 | — | — |
| AIME 2026 | 94.7 | — | — |
| OSWorld-Verified (computer use) | 65.9 | — | 75.6 🥇 |
| TerminalBench 2.1 | — | — | 60.7 🥇 |
Benchmarks from MarkTechPost's full benchmark report. Muse Glimmer leads on agentic orchestration (MCP Atlas). Qwen3.6-27B leads on computer use and terminal execution. Different models for different primary workflows.
How to Download and Run It
Ollama (easiest): ollama run meta/muse-glimmer — downloads and runs automatically on Mac or Linux. Requires 24GB VRAM or Apple Silicon with 32GB unified memory.
llama.cpp: Download Q4_K_M GGUF from Hugging Face (meta-ai/Muse-Glimmer-GGUF). Fits in 24GB VRAM. 3.1× speed with DFlash speculative decoding drafter included in the release.
LM Studio / vLLM / SGLang: All supported at launch. vLLM and SGLang for server-side deployment at scale. LM Studio for desktop GUI.
Security note: Like NVIDIA NOOA, Muse Glimmer agents that execute code or call external tools should run inside a container or VM. Meta's own safety evaluation shows a 28.4% attack-success rate on Siren AgentDojo (prompt injection). Do not deploy Muse Glimmer agents with sensitive data access outside an isolated execution environment.
Who Should Use It
Teams building local agents: The first Apache 2.0 30B agentic model that fits on hardware most developers already own. MCP Atlas #1 among local models. No per-token charges, no data to third-party APIs, works offline. Fine-tune on your own data without licensing restrictions.
LLM-as-a-judge pipelines: Meta explicitly targets this use case. Local judge model that evaluates outputs without sending data to OpenAI or Anthropic.
Not for: Computer-use agents (Qwen3.6-27B leads on OSWorld-Verified). Terminal execution agents (Qwen leads TerminalBench). Production regulated workloads where Anthropic FLI C+ safety posture is required. Anything where 28.4% prompt-injection attack success is unacceptable without mitigations.
Sources: Meta official model page · AI Weekly · MarkTechPost benchmarks · VentureBeat · Neowin · Phoronix · Related: NVIDIA NOOA review → · AI model release tracker →