GitHub radar
Meta Muse Glimmer 30B: Agentic Local AI
Meta Superintelligence Lab released an open multimodal model for local agentic tasks — it sees images, calls tools, and executes multi-step workflows without a cloud.
Muse Glimmer 30B is an official open model from Meta Superintelligence Lab, released in August 2026. It is a multimodal language model with 29.8 billion parameters that understands both text and images, calls external tools, and executes multi-step tasks autonomously. The architecture combines Gated DeltaNet and Gated Attention across 64 layers, with a built-in 1.8B-parameter vision encoder. Native context is 131K tokens (extendable to 1M). DFlash 2 speculative decoding delivers a 3.1× inference speedup. Licensed Apache 2.0 for commercial use. Benchmark highlights: τ3-Bench and MCP-Atlas (agentic), SWE-Bench (engineering coding), DeepSearch QA (multi-step retrieval).
Why a vibe-coder should care
This is the first serious agentic model from Meta that can run on consumer hardware without a cloud. It replaces the combination of a separate vision model, LLM, and tool layer — everything is built in and works as a single agent. Useful for automating tasks that require reading screens, making decisions, and calling external services.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Meta's Muse Glimmer 30B locally: https://huggingface.co/meta-models/Muse-Glimmer-30B — find a quantized version for my Mac, run it via llama.cpp or transformers, and show me a multimodal example with an image.
A Mac with 32 GB of RAM or more, or a GPU with 24 GB — the 4-bit version takes around 18 GB.
Open on Hugging Face▌ More finds