Meta released Muse Glimmer on August 10, 2026 — a 30-billion-parameter open-weight model distilled from Muse Spark and built specifically for local agentic workloads. It is available on Hugging Face under Apache 2.0, it runs on a consumer GPU or Apple Silicon Mac, and two Ollama commands get you from zero to a working local agent. No API key. No per-token billing. No data leaving your machine.
If you have been watching cloud AI costs climb as your agentic pipelines get more autonomous, Glimmer is worth a closer look.
Why Apache 2.0 Is the Real Headline #
The license is as important as the model weights. Apache 2.0 means no commercial restrictions, no ambiguous acceptable-use policies, and no dependency on Meta’s infrastructure. Muse Spark — Meta’s frontier model — remains closed and API-only. Glimmer is the version developers actually own. You can fine-tune it, ship it in a product, run it air-gapped, and the license does not change on you.
This puts Glimmer alongside Qwen3.6-27B as one of the most permissive capable models in the 27–31B weight class. For teams with data governance requirements or enterprise customers who push back on third-party API calls, that matters more than a few benchmark points.
Hardware Requirements #
The 4-bit K-Quant GGUF weighs 16.76 GB and runs on 18 GB or more of VRAM or unified memory. In practice, a 24 GB GPU gives you comfortable headroom for the KV cache and runtime buffers. The MLX variant works well on Apple Silicon Macs with 24 GB or more of unified memory — M2 Pro, M3 Max, M4 Pro, and above.
AMD’s Ryzen AI Max platform is also officially supported, with AMD publishing a setup guide at launch. For most developers, full-precision inference (requiring 64 GB+) is unnecessary. The K-Quant variant is what you want.
Getting Started with Ollama #
If you already have Ollama installed, this is your entire setup process:
ollama pull muse-glimmer:30b
ollama run muse-glimmer:30b
ollama pull muse-glimmer:30b-mlx
LM Studio has the model in its library as well. For lower-level control or llama.cpp integration, the GGUF files are on Hugging Face at meta-models/Muse-Glimmer-30B-GGUF and unsloth/Muse-Glimmer-30B-GGUF. For OpenAI-compatible local API serving, vLLM and SGLang both work out of the box. If you need a cloud fallback without committing to a full local setup, Together AI, Fireworks, and OpenRouter already carry the model.
Built for Agents, Not Chat #
Glimmer was not optimized to win general-purpose chat benchmarks. Meta trained it specifically for agentic tasks: tool use, function calling, multi-step reasoning, multimodal input, and failure recovery. The benchmarks reflect this specialization. On AIME 2026 it scores 94.7% — strong for a 30B model. On SWE-Bench Pro it lands at 51.2%, which is competitive in its weight class. Meta also benchmarked it on MCP-Atlas, τ-Bench, and DeepSearch QA — tasks that measure how well a model operates inside scaffolds, handles tool calls, and completes multi-turn agentic requests.
These are not synthetic toy tasks. They measure what the model actually does in production agent environments. The comparison set is Gemma4-31B and Qwen3.6-27B — both solid local models. Glimmer holds its own.
That said, be clear-eyed about where a 30B model fits. Complex frontier-scale reasoning, tasks that require the highest possible accuracy, and cutting-edge multimodal work still belong to cloud models. Glimmer is not trying to replace GPT-5 or Claude Opus 5. It is trying to replace the API call you make for routine agentic tasks — and at that job, it does well.
Glimmer vs Spark: When to Use Each #
Meta’s lineup now divides cleanly. Muse Spark is the closed frontier model — hosted, billed per token, no offline option, best accuracy. Muse Glimmer is the open local complement — Apache 2.0, runs on your hardware, no variable costs, built for always-on agentic workloads.
The practical decision is straightforward. Use Glimmer if you need an always-on local agent, if data governance or privacy requirements apply, if your workload involves frequent calls where per-token billing adds up, or if you are experimenting and do not want to watch a spending dashboard. Use Spark if the task is hard enough to require frontier capability.
For many real-world agentic pipelines, Glimmer handles the majority of the work — routine tool calls, document analysis, code review, context summarization — and Spark handles the exceptions. That hybrid approach is already showing up in how the developer community is deploying both models.
Where to Start #
The model weights are available now at meta-models/Muse-Glimmer-30B on Hugging Face under Apache 2.0. The official announcement is on the Meta AI Research blog. For hardware-specific setup, the NVIDIA developer blog and AMD have both published integration guides. If you want to understand the architecture in more detail, the Meta developer documentation covers the model design and benchmark methodology.
Ollama is the fastest path from announcement to running agent. If your machine qualifies, there is no good reason not to try it today.