Meta's Muse Glimmer: a 30B Apache-2.0 vision agent you run locally
Meta's first big open-weights model in a while runs on consumer GPUs, posts strong agent and coding scores — and GLM-5.2 is free on Vercel through Aug 27.
Copy markdown
Meta is back in open weights, Apache-2.0
Muse Glimmer 30B is Meta's first major open-weights model in a while, under a permissive Apache-2.0 license — commercial use, no login gate. It pairs a 2B vision encoder with a 28B text decoder and ships native tool-calling and object detection, built for local agents rather than chat.
Runs on hardware you already own
Day-0 support in transformers, llama.cpp, and vLLM, plus GGUF quants that fit a single 24GB consumer GPU — one widely-shared demo runs the whole model in memory on a 16GB Mac mini. A capable local vision agent is now a weekend build with no API bill.
The scores lean agentic, not just chat
It posts 75.5 on MCP Atlas (tool use) against 62.5 for Qwen 3.6 and 54.2 for Gemma 4, plus 51.2 on SWE-Bench Pro and 94.7 on AIME 2026 — strong numbers for a 30B you can run offline.
Elsewhere: GLM-5.2 is free on Vercel to Aug 27
Z.ai's open-weights coding model, with a usable 1M-token context, is free via Blackbox on Vercel's AI Gateway through Aug 27 — a no-cost way to point an agent at a whole codebase this week.