MOSS-Image: Open 5B Model Renders Long Text with 96.3% Coverage

The MOSS-Image team has released a technical report on an open image generation system that targets the weakest spot of open-weight models: dense text and complex layouts. The system is built on a 5B-parameter diffusion transformer trained from scratch. An agentic module sits on top and turns a short request into a detailed description of content and layout. On the new LongLongTextBench benchmark, average text coverage reached 96.3%. That is the best result among models with open weights.

Where open generators break down

Closed systems can already produce more than landscapes and portraits. They also handle information-dense graphics: posters, slides, annotated diagrams. Open text-to-image models lag behind here. Ask for a poster with a paragraph of small print or an infographic with several blocks, and the letters blur, lines disappear and elements overlap. The authors of MOSS-Image state their goal plainly: to close this gap.

The approach has two layers. The first is the model itself, a single-stream DiT with 5B parameters. It was pretrained from scratch with a purpose-built recipe rather than fine-tuned from someone else's base. The result is bilingual rendering of long text in English and Chinese, along with precise layout control.

The second layer is an intention-aware agentic prompt enhancement module. A user writes "concert poster", and the module identifies the intent and expands the request into a specification: which captions are needed, where they go, how the blocks are arranged in a hierarchy. The generator receives not a vague phrase but something close to a finished technical brief.

Two new benchmarks and what they showed

The authors argue that existing tests do a poor job of checking long text and end-to-end layout generation, so they added two of their own:

• LongLongTextBench (LLTB) evaluates long-text rendering in English and Chinese; MOSS-Image reached an average text coverage of 96.3%, the best result among open-weight models.
• RealLayoutArena (RLA) evaluates end-to-end generation of layout-heavy images; in the comparison of full systems, MOSS-Image placed second.
• The backbone model on RLA ranked first among generators evaluated with the same expanded prompts.
• Model size is 5B parameters, with a single-stream DiT architecture.

The last two points matter more than the rest. They show that the result does not come down to a lucky prompt-writing agent. When all generators receive identical expanded prompts, the MOSS-Image backbone is still ahead. Second place in the end-to-end comparison means the full system currently trails someone on this test, and the abstract does not say who.

How to read this

For practitioners, the takeaway is this: the model deserves a place in the queue for evaluation, but it is not something to plan production around right now. The authors promise to release the model, the training recipe and the evaluation framework, and the abstract phrases this in the future tense. Until the weights are publicly available, the numbers remain the authors' own claims.

Both benchmarks were created by the same team, and 96.3% text coverage does not equal 96.3% error-free images. Coverage shows what share of the requested text appeared in the picture; spelling accuracy, fonts and composition will have to be checked separately. Judging by the description, the agentic module adds a request-processing step before generation, so the latency and cost of the whole chain need to be measured.

The best way to test it after release is to take your own tasks, such as ad banners, slides, covers and infographics in two languages, and run them on the model with and without the agentic module. If the text quality holds up, teams that need control over the weights and the ability to fine-tune the model will get an open alternative to closed APIs. The training recipe and evaluation code may prove more valuable than the weights themselves: they would let other labs reproduce the result and compare their own models on the same tests.