Meta opens Muse Glimmer, an agentic 30B model built for consumer GPUs

Muse Glimmer is Meta's new 30-billion-parameter open model built specifically for local, always-on agentic workflows. Released under an Apache 2.0 license, it allows developers to run private, tool-using assistants entirely offline on a single consumer GPU or Mac.

What it actually does

Instead of acting like a basic chatbot, Muse Glimmer is trained to plan steps, invoke functions, and handle long workflows on your personal computer.

It features a 2B perception encoder that accepts interleaved text and images, allowing it to interpret screenshots, charts, and local files directly alongside conversations.

When a function call breaks or returns bad data, the model does not just halt. It is trained to diagnose the error and retry alternative steps automatically.

It also works with agent frameworks like OpenClaw, supports adjustable reasoning effort depending on whether you prioritize speed or quality, and handles over 100 languages. You can point it at personal scheduling, local coding, or document analysis without sending a single byte to the cloud.

How it fits on consumer hardware

At full precision, a 30B parameter model typically requires over 55GB of memory, which is completely out of reach for consumer setups.

Meta used quantization techniques to compress the weights down to roughly 4-bit precision, shrinking the model to under 20GB. A K-Quant-17GB release leaves enough room for the KV cache, the vision encoder, and a speculative decoding drafter within a standard 24GB or 32GB memory envelope.

To keep local generation feeling responsive during long reasoning chains, it ships with an optional DFlash-based companion drafter. This model proposes blocks of tokens in parallel, which the main model quickly verifies, significantly accelerating code and text generation.

Hardware like an NVIDIA RTX 5090 or Apple silicon M4 Max and M5 Max chips can run it smoothly. The weights are live on Hugging Face, with ecosystem support landing across llama.cpp, MLX, vLLM, and Ollama.

Benchmarks and mixed reactions

On Meta's published benchmarks, Muse Glimmer scores 75.5 on MCP Atlas and 51.2 on SWE-Bench Pro, beating similar-sized models like Gemma4-31B across most agent-heavy tasks.

However, community reviewers have pushed back against some of Meta's framing. Reviewers noted that Qwen3.6-27B actually outperforms Muse Glimmer on several key benchmarks despite having fewer parameters.

For example, Qwen3.6 scored 75.6 on OSWorld-Verified compared to Glimmer's 65.9, and also edged it out on TerminalBench 2.1. Outside evaluations show the model is strong at agentic tooling, but not an uncontested champion across the board.

Why it matters

If you are a developer or power user who cares deeply about privacy, cost, or offline autonomy, Muse Glimmer is worth spinning up immediately. The permissive Apache 2.0 license makes it completely safe for commercial applications and custom hacking.

Packing multimodal understanding, resilient tool calling, and speculative decoding into under 20GB lowers the barrier for running autonomous agents without cloud dependencies.

We think it is a smart release that directly serves the local AI community, even if it does not blow away rivals like Qwen on every raw task. If you have a machine with 24GB of VRAM or unified memory, it is easily one of the most capable local agent foundations you can test today.