MOSS-Image Is Out: OpenMOSS Shipped a 5B DiT That Takes Bounding Boxes in the Prompt

The awkward detail behind MOSS-Image's launch is that the loudest thing said about it is not quite true yet. On October 17, vLLM's own account congratulated the OpenMOSS team on the open-source release and said it was "proud to support MOSS-Image on Day 0 with vLLM-Omni." The pull request that actually adds the model to that runtime — [Model] Add MOSImage text-to-image with SLG, 3,605 added lines across 25 files — was opened on October 14 and is still open, with the reviewer asking for output-parity evidence against a pinned reference implementation. vLLM-Omni's newest published release is v0.31.0rc1, dated October 6. So no shipped version of vLLM-Omni contains this model, and the honest version of the Day 0 claim is that the serving path exists as a branch you can build, not as a package you can install. That distinction matters more than usual here, because MOSS-Image is a genuinely interesting release whose real content is in the weights: MOSS-Image-Base and MOSS-Image-SFT, released on October 16, 2026 under Apache 2.0, built around a compact diffusion transformer that accepts structured JSON layouts as a prompt.

The short form for someone who wants the shape of it. OpenMOSS — the lab behind the MOSS language and speech line, working with MOSI Intelligence and academic partners — shipped a 5B single-stream diffusion transformer trained from scratch, paired with a frozen Qwen3.5-9B text encoder and a frozen FLUX.2 VAE, under a permissive licence that covers both weights and inference code. Its selling point is not raw image quality; it is that you can hand it a JSON object containing bounding boxes, literal text strings and typography notes, and it will honour them. The benchmark numbers attached to it are the vendor's own, measured on two benchmarks the vendor also built. That is not disqualifying, but it does mean you should read them as a lab's internal scoreboard rather than as a league table.

What actually shipped, and where

The artifacts are straightforward and checkable. On Hugging Face, OpenMOSS-Team/MOSS-Image-Base and OpenMOSS-Team/MOSS-Image-SFT carry the weights, with ModelScope mirrors for both, a Gradio Space running on a single A10G, and a GitHub repository holding the pipeline, the ComfyUI plugin and a technical report. The project page at openmoss.ai stamps the same date: October 16, 2026 — Introducing MOSS-Image-Base, with code and model weights available.

Two checkpoints, one architecture. Base is the pretrained model; SFT is fine-tuned on curated data. The vendor's own comparison example shows SFT producing tighter text grouping and a more organised layout, which is roughly what you would expect from a curated-data fine-tune of a base model that was deliberately left undistilled. The Base card is explicit that it "has not undergone any post-training or distillation of inference parameters such as CFG", the stated reason being to preserve training information and make the model easier to train LoRAs on. The practical consequence is that Base is the better host for fine-tuning and the worse default for casual prompting.

Licensing is Apache 2.0 for the weights and for the standalone pipeline code — no revenue threshold, no acceptable-use annex, no separate commercial tier. For a model whose headline capability is rendering legible text into images, that is a more consequential fact than the benchmark table, because text rendering is exactly the capability that shows up in commercial work: product mockups, UI comps, infographics, posters, ad layouts.

The architecture, in the numbers that matter

The technical report's configuration table is where the release stops being a press release and starts being a spec. The 5B figure refers to the diffusion generator alone — a single-stream transformer, 42 joint blocks deep, hidden width 3072, 24 attention heads at 128 dimensions each, an 8192-wide feed-forward using SiTU-GLU activations, and a four-axis rotary position encoding whose first axis is reserved for a reference index, so reference-conditioned generation can be added later without changing the positional scheme.

Then there is the part the "5B" label quietly excludes:

• Text encoder — Qwen3.5-9B, frozen, feeding final-layer token hidden states at dimension 4096 through a linear projection to 3072 plus two text-only transformer blocks. A 9B encoder bolted to a 5B generator is a materially bigger footprint than the headline number suggests.

• VAE — FLUX.2's VAE, frozen and loaded from FLUX.2-klein-4B, 32 latent channels at a spatial downsampling factor of 8, with 2×2 latent patches giving an effective image stride of 16 pixels. MOSS-Image's own model card notes it uses the FLUX.2 small decoder.

• Download size — the SFT repository totals about 27.2 GB, of which the text encoder is roughly 16.8 GB across four shards and the transformer around 10.2 GB across three. Plan for the encoder, not the badge.

• Prompt ceiling — 4096 tokens at the final pretraining stage, matching the vendor's "up to 4K tokens" claim. Output dimensions must be positive multiples of 16, with the guide recommending one, one and a half, or two megapixels and a supported extrapolation range from 0.25M (512×512) to 4M.

• Sampling defaults — flow-matching Euler, 50 steps, CFG 3.0, shift 5.0, with skip-layer guidance on the negative branch at layer 22. The default guidance path needs two transformer forwards per step, which is why 50 steps is not the cheap setting it sounds like.

• Training — three progressive resolution stages: 256px for 2.1M steps, 512px for 2.1M steps, then roughly 1024² for 0.31M steps, with the maximum tokens per sample rising from 576 to 3,072 to 8,192. The corpus is a ReLAION-derived set trimmed from about 5 billion to 3 billion images by deduplication and AI-image removal, with the report stating that no further filtering is applied — images whose captions contain artefacts are retained and their artefacts described, rather than dropped. Optimisation is a hybrid Muon–AdamW, and the vendor's own 1B-parameter ablation reports 33.2% and 28.6% fewer steps to the same loss than AdamW at two checkpoints. Training is stated to have used heterogeneous GPU platforms, which is a claim about hardware supply rather than about the model, and the report does not name them.

One thing worth flagging before the benchmarks

The prompt enhancement module — the piece that turns "design an English koala knowledge infographic" into a full content and layout specification — is not a small model of the vendor's own. The report states plainly that LayoutPE uses G​PT-6 with modular task instructions at inference time, and the repository's bundled prompt-enhancer skill is described as being designed for G​PT-6-family agents. So the strongest end-to-end result below is not a single model's output; it is a two-model pipeline whose specification half runs on someone else's API. If you were reading the Elo ranking as a pure measure of the generator, it is not that.

The benchmark table, and who authored it

MOSS-Image is evaluated on two benchmarks that ship with the release, born of the vendor's own argument that existing tests do not measure the thing it cares about. LongLongTextBench is 240 literary-text rendering tasks — 120 English, 120 Chinese, at a mean target length of 1,339.8 characters, roughly eleven times LongTextBench's average. RealLayoutArena is 61 design tasks (20 posters, 20 infographics, 21 user interfaces) judged by blind pairwise human comparison and fitted to a Bradley–Terry Elo scale. Where baselines appear, they are the vendor's runs of other people's models. Every figure in this section is vendor-reported, unreproduced, and none of it has been independently audited.

On LongLongTextBench's Mean-6 text coverage, MOSS-Image scores 96.3 — the best open-weight figure in the vendor's table, well clear of Q​wen-Image-2512's 64.9, and its longest-condition scores are 97.1 English and 92.4 Chinese. The number that keeps the table honest is the row above it: on the vendor's own benchmark, on the vendor's own runs, GPT Image 2 scores 99.0 and Nano Banana Pro 96.6. OpenMOSS published a table in which a competitor still wins, which is worth something.

On 7Bench layout control, the gap is much wider — mean IoU over detected target objects is 0.88 for MOSS-Image against 0.44 for I​deogram 4.0, 0.33 for Nano Banana Pro and Seedream 5.0 Pro, 0.31 for Q​wen Image 3.0 Pro and 0.30 for GPT Image 2, with a Layout AUC of 0.70 against a field running 0.16 to 0.20. If those numbers survive reproduction, precise object placement is the capability this model actually owns.

On RealLayoutArena end-to-end, MOSS-Image lands at 1698.70 Elo with a 95% confidence interval of 1655.96 to 1741.43 — second overall, behind GPT Image 2 at 1844.35 and ahead of Seedream 5.0 Pro (1661.16), Q​wen Image 3.0 Pro with prompt enhancement (1614.84) and Nano Banana Pro (1590.19). When the comparison is restricted to the backbone alone under shared enhanced natural-language prompts, MOSS-Image comes first at 1807.34, ahead of ERNIE Image at 1748.51 and SenseNova-U1 at 1689.23. The prompt-strategy ablation is the most useful line in the paper for anyone planning to use it: JSON layout prompts score 1622.87, natural-language prompts 1506.20 on the same system. The layout interface is not a bonus feature. It is where the performance lives.

Two caveats the report raises against itself are worth repeating. LongTextBench is described as close to saturation for strong current models, which is the vendor explaining why it built a harder test rather than resting on an old one. And the confidence interval on the end-to-end Elo overlaps nothing, but it is wide — 85 Elo points — on a 61-task arena judged by human raters.

Running it yourself, and what that costs

Today there is no commercial API for MOSS-Image that we can verify, and it is not on our own catalogue — OrcaRouter serves 200-plus models and MOSS-Image is not one of them, so nothing here should be read as us offering it. What exists is self-hosting: the Diffusers route with a custom pipeline (pipeline_moss_image, diffusers 0.38 or later, transformers 5.3 or later, torch 2.9 or later, Python 3.10+), a Gradio app, and a native ComfyUI plugin installable through ComfyUI Manager, with integration into the official ComfyUI repository described as in progress. There is also a prompt-enhancer skill for agent frameworks, and it compiles pixel-coordinate drafts into the model's centred-coordinate JSON format without calling any image service — a small, genuinely useful piece of tooling.

The community has already done the part the vendor did not. Within roughly a day of release, third-party builds appeared on Hugging Face: a Comfy-Org packaging of the model, an experimental Kijai build, GGUF quantisations, INT8 and INT4 ConvRot conversions, a W4A8 OrbitQuant build, and — most useful for anyone without a datacentre — a 4-step DMD LoRA and an 8-step distilled variant that attack the 50-step default head-on. None of these are vendor-supported and none are benchmarked, but they are the normal signal that a release has landed with people who intend to use it rather than cite it.

This is where a routing layer earns its place, and it is worth being precise about the boundary. We do not host MOSS-Image, so we cannot serve it and will not imply otherwise. What we can do is the other half of the comparison: if you want to measure a self-hosted MOSS-Image against hosted image models, the hosted side is reachable through one API key on our catalogue at provider list price with 0% markup — so a vendor price change on a hosted model is live here the same day, not on a quarterly contract cycle — with automatic failover across providers and a routing DSL that lets you place several models behind a single endpoint. Testing an unproven checkpoint against a rented one should not require a second SDK, and that is the only claim being made.

Who should move on this, and who should wait

If your work involves generating images that contain correct, legible, multilingual text in specific positions — posters, interface mockups, information cards, anything where the layout is the deliverable rather than a happy accident — MOSS-Image is the most directly targeted open-weights release of the last month, and the Apache 2.0 licence means the usual commercial review is short. Start with the SFT checkpoint, expect to supply structured JSON rather than prose, and budget for the encoder's 16.8 GB before you budget for the generator.

Wait if you need a callable endpoint this week, if you are renting capacity rather than owning it, or if you want an independent number before you commit. The architecture is a clean single-stream design with a frozen encoder and a frozen VAE, which will make it straightforward to run — but 50 steps at two forwards per step is a real serving cost, the "5B" badge understates the download, and the headline scores were produced by the vendor on the vendor's own benchmarks using a specification stage that runs on a third-party model. Wait for a reproduction, or for the vLLM-Omni path to merge into something you can install.

Two questions this release does not settle

Does the model's layout control justify a workflow change?

That depends on whether you can express your intent as boxes and strings. The layout JSON uses absolute pixel coordinates measured from the canvas centre, with bbox_format "yxyx", which is unlike the relative-coordinate conventions several other recent layout systems use — the vendor argues absolute pixels suit multi-resolution work better. Adopting it means rewriting whatever prompt logic you already have, and the vendor's own 7Bench gap (0.88 mIoU against a field at 0.30–0.44) is the only evidence that the rewrite pays. It was measured by the people who wrote the metric.

Is the Fine-Tuning story as open as the licence implies?

Partly unknown. The weights and pipeline code are Apache 2.0 and Base was deliberately left undistilled so LoRAs would have something to bite on, which is a real commitment. But the report's training recipe — the staged resolution curriculum, the ReLAION filtering decisions, the Muon–AdamW split — is documentation, not released data, and the fine-tuning guidance so far consists of specific instructions (keep the text encoder frozen; do not fine-tune it) plus a captioning prompt. The claim that simple SFT unlocks reference-based generation and GIF output is stated as a finding with a "stay tuned" attached, not as a result you can check today.

Three things would move MOSS-Image from an impressive repository to a settled choice, and none of them exist yet: an independent run of LongLongTextBench or RealLayoutArena by anyone other than the authors, the vLLM-Omni patch merged into a tagged release with the output-parity evidence its reviewer asked for, and a few-step distilled checkpoint with a vendor blessing rather than only community conversions. Until then the accurate summary is this — a well-documented 5B diffusion transformer with an unusually strong text-and-layout interface, a permissive licence that covers the code as well as the weights, benchmark leadership on tests it wrote itself and second place on the end-to-end comparison, and a serving story that is real but still sitting in an open pull request.

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.