CogView-4 is Zhipu AI's open-weight text-to-image model, released on March 4, 2025 — the dated suffix in the model slug, 250304, is that release date, not a version number. It was the first open-source text-to-image model to render Chinese characters natively and accurately inside generated images, and the first image generation model of any real capability to ship under the Apache 2.0 license rather than a research-only or non-commercial one. Both firsts matter for the same reason: most open image models before it were either weak on non-Latin script or licensed in a way that kept them out of production.
Where it shines
Text rendering inside images is the standout feature, specifically for Chinese: signage, packaging mockups, posters, and UI screenshots with embedded Chinese copy are the use case the model was built to solve, and it is measurably better at this than the open models that preceded it. It is bilingual — Chinese and English prompts both work, with prompt length up to 1024 tokens, which is enough room for detailed, multi-clause scene descriptions rather than short keyword prompts.
Under the hood
CogView-4 is a 6-billion-parameter diffusion transformer using GLM-4-9B as its text encoder — the same text-understanding backbone Zhipu uses across its language model line, which is part of why it handles longer and more structured prompts well. It supports resolutions from 512 up to 2048 pixels on a side, constrained so that width times height stays under roughly two million pixels, giving flexibility across square, portrait, and landscape outputs rather than a fixed set of presets.
Prompting notes
Longer, more explicit prompts are rewarded here rather than penalized — the GLM-4-9B text encoder is built to parse structured, multi-part descriptions. For Chinese-text-in-image work specifically, putting the exact characters you want rendered directly in quotes within the prompt gets more reliable placement than describing the text indirectly.
Where it falls short
Zhipu's own documentation is candid that Chinese text accuracy, while a genuine advance, is still imperfect — expect occasional stroke errors or malformed characters, especially at smaller text sizes or busier layouts. Self-hosting is not casual: the model requires at least 32GB of system RAM to avoid the process being killed, and generating at the top resolution needs roughly 39GB of GPU memory without offloading, which puts it well outside consumer-GPU territory. Inference precision is limited to BF16 or FP32 — no lower-precision quantized path is offered by the maker.
Rights and licensing
CogView-4 ships under the Apache 2.0 license — permissive for commercial use, with no field-of-use restriction beyond standard attribution. Zhipu has published it as the first image model of its caliber to use these terms rather than a research or non-commercial license, which is the reason it turns up in production pipelines rather than only in research demos.
When to pick it
CogView-4 is the right pick when embedded Chinese text is a real requirement — packaging, signage, marketing creative for Chinese-language markets — and when a permissive open license matters more than having the single best general photoreal model available. It's also a reasonable choice anywhere a team wants to self-host rather than depend on a hosted API, provided the GPU budget is there.
Alternatives worth comparing
For even stronger embedded-text accuracy at the cost of much heavier compute, Zhipu's newer glm-image targets the same text-rendering problem with a larger hybrid architecture. For general-purpose photoreal or stylized generation without a text-rendering requirement, gpt-image-1 and gemini-2.5-flash-image are the more broadly benchmarked hosted options.
Last technical review: 2026-09-14 — Tokonomix.ai