Z.ai just dropped GLM-Image, an open-source powerhouse blending a 9B-param auto-regressive generator (spun from GLM-4) with a 7B diffusion decoder. The AR side cranks out semantic-VQ tokens for solid layouts and precise text placement, dodging the "semantic drift" that trips up pure diffusion models like Flux or SD3.5. It shines in benchmarks: tops CVTG-2k word accuracy at 91.16% vs Nano Banana's 77.88%, crushes LongText-Bench for long prompts, and holds strong on DPG Bench for dense info.
Trained progressively from 256px to 2048px, it supports text-to-image, editing, style transfer, and more under MIT license for easy commercial tweaks. What sets it apart? Killer text rendering in infographics or diagrams, where it reasons like an LLM before visualizing, beating Qwen-Image-2512 in spots. Downsides: hands or photoreal scenes can look off in demos, and it's not always the prettiest artistically.
Still, at low cost via APIs, it's a solid pick for enterprise needs like UI mocks or multilingual ads, pushing open models closer to closed ones without the price tag.
