Meta's Muse Glimmer: an open 30B agentic model on one GPU
Apache 2.0, multimodal, and it fits a 32GB card or an M4-Max Mac — a private local agent that reads your screen and calls your tools. GGUF quants are live.
Copy markdown
Apache 2.0 — you can actually ship it
The full 30B (a 2B perception encoder plus a 28B decoder) ships under Apache 2.0: fine-tune it, embed it in a product, no gated license or usage caps. Meta says it beats Qwen3.6-27B and Gemma4-31B on agentic, coding and multimodal tasks.
Runs on one 32GB card — or your Mac
GGUF K-quants land around 17–20GB, so it fits a 24–32GB GPU or an M4/M5-Max MacBook; full BF16 wants ~64GB. Serve it via llama.cpp, vLLM, SGLang or NVIDIA NIM, and a built-in DFlash speculative drafter gives ~3.1x on an RTX 5090.
Built to be an agent, not a chat toy
Strict-schema tool calling, multi-step reasoning with failure recovery, and a 'controllable effort' dial to trade speed for quality. It posts 51.2 on SWE-Bench Pro, 75.5 on MCP Atlas and 74.6 on DeepSearch QA — strong numbers for a model you host yourself.
It sees your screen — build a local desktop agent
Interleaved text and image input reads screenshots, charts and PDFs (75.4 ScreenSpot Pro, 78.8 CharXiv) across 100+ languages. This weekend: wire it to your tools for a fully private agent whose data never leaves your machine.