MOSS-Image vs MAI-Image-2.6 Preview: Both Ranked Second, Neither One You Can Call

M​OSS-Image and M​AI-Image-2.6 Preview have more in common than the fact that both are new image models. M​OSS-Image was released on October 16, 2026 as open weights — the M​OSS-Image-Base and M​OSS-Image-SFT checkpoints under Apache 2.0. M​AI-Image-2.6 Preview entered private preview on August 19, 2026 on M​icrosoft's M​AI Playground and M​icrosoft Foundry. Each one currently sits at number two on a leaderboard, the two leaderboards were built by different people using different methods, and neither model has a price you can look up.

That last point is the whole comparison. This is not a head-to-head between two things you can rent and measure against each other. It is a matchup between a 27 GB download from a research lab and an invite-gated preview from a platform company, and the useful question is not which one scores higher — it is which one you can actually get into a product this quarter, and what each vendor is willing to tell you before you commit.

Two second places, and what each one is measuring

The numbers that get quoted for both models are both "No. 2", and that coincidence hides the fact that they measure almost opposite things.

M​OSS-Image's second place comes from RealLayoutArena, a 61-task design benchmark — posters, infographics and user interfaces — that ships with the model's own technical report. It scores 1,698.70 Elo with a 95% confidence interval of 1,655.96 to 1,741.43, behind GPT Image 2 at 1,844.35. RealLayoutArena was built by OpenMOSS, the lab that built M​OSS-Image, and the baselines in it are the lab's own runs of other companies' models. It is a vendor scoreboard, and the vendor has labelled it as one.

M​AI-Image-2.6 Preview's second place comes from Arena, the rebranded LMArena, where blind human voters pick the better of two images from the same prompt. The entry is literally named mai-image-2.6-preview and reads 1,336 Elo with a ±11 margin on roughly 3,488 votes — behind GPT Image 2 Medium at 1,381 and ahead of G​rok Imagine Image 2.0 at 1,316. The same preview build later landed at No. 3 on Arena's single-image-edit board at 1,420 points. That is an independent human-preference measure, and it is also a preliminary one: the top entry on that board carries more than 70,000 votes, so the margin of error on a 3,488-vote entry is real, and early-vote Elo moves.

So the honest framing is that one model is second on a test its own authors wrote, and the other is second on a crowd board with a fifth of the evidence behind it. Neither is a settled result, and neither has been reproduced by anyone outside the organisation that published it.

What M​OSS-Image actually gives you

M​OSS-Image is a 5B single-stream diffusion transformer trained from scratch: 42 joint blocks, hidden width 3,072, 24 attention heads at 128 dimensions, an 8,192-wide SiTU-GLU feed-forward and a four-axis rotary position encoding. The headline parameter count excludes a lot. The text encoder is a frozen Qwen3.5-9B, and the VAE is FLUX.2's, loaded frozen from FLUX.2-klein-4B, with 32 latent channels and a 2×2 latent patch on top of an 8× downsample for an effective stride of 16 pixels. The SFT repository is about 27.2 GB, of which roughly 16.8 GB is the encoder.

Its defining capability is not image quality. It is that the prompt can be a JSON object containing bounding boxes, literal text strings and typography notes, expressed in absolute pixel coordinates measured from the canvas centre with bbox_format "yxyx". The vendor's own ablation makes the case bluntly: on RealLayoutArena, JSON layout prompts score 1,622.87 while natural-language prompts on the same system score 1,506.20. The structured interface is not a bonus feature bolted onto the model. It is where the performance lives.

The supporting evidence is a benchmark the lab also built, 7Bench, where M​OSS-Image reports a mean layout IoU of 0.88 against 0.44 for I​deogram 4.0, 0.33 for N​ano 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. On the text-rendering side it reports 96.3 Mean-6 on LongLongTextBench — a 240-task literary-text benchmark it also authored — with 97.1 in English and 92.4 in Chinese. The most credible line in that table is the one above it: on the vendor's own benchmark, GPT Image 2 scores 99.0 and N​ano Banana Pro 96.6. Publishing a table where a competitor still wins is worth something.

What M​OSS-Image does not give you is an endpoint. There is no commercial API we can verify, no hosted playground beyond the Gradio Space running on a single A10G, and no managed path. You install Diffusers with a custom pipeline, or the native ComfyUI plugin, and you supply the GPU.

What M​AI-Image-2.6 Preview actually gives you

M​icrosoft has published far less about its model than OpenMOSS has about its own, and the pattern of what is missing is informative.

What exists: a family lineage that moves on a six-to-eight-week cadence — M​AI-Image-1, then M​AI-Image-2 in March 2026 with the photorealism and text-rendering jump, M​AI-Image-2.5 in June 2026, which became the engine inside Bing Image Creator, PowerPoint and OneDrive, then M​AI-Image-2.5-Pro in July 2026 with an 8K premium tier, and now 2.6 roughly six weeks later. M​icrosoft's own capability descriptions are qualitative: stronger text rendering, portraits and 3D imagery; more polished commercial and photorealistic output; work across multiple reference images in a single generation; richer grounding; greater control over reasoning, format and resolution. No parameter count, no licence, no architecture, no benchmark table.

What does not exist: a price. The preview is invite-gated, the access route is M​AI Playground and M​icrosoft Foundry, and there is no per-image rate published. There is also no general-availability date. Arena's category breakdown is the only granular result attached to the model — it ranks first in 3D imaging and modelling, and second in cartoon/anime/fantasy, product and branding, text rendering and art — and category ranks on a low-vote preliminary board are directional at best.

The one thing that is clearly true and clearly significant is the cadence. M​icrosoft is shipping image-model improvements faster than this market is used to, and the Arena debut suggests the compounding is real rather than announced.

The two models, line by line

• Release — M​OSS-Image on October 16, 2026 as open weights; M​AI-Image-2.6 Preview into private preview on August 19, 2026

• Access — M​OSS-Image downloadable and self-hosted, with a ComfyUI plugin and a Diffusers pipeline; M​AI-Image-2.6 Preview invite-gated on M​AI Playground and M​icrosoft Foundry

• Price — M​OSS-Image free to download, with a ~27.2 GB footprint and your own GPU time; M​AI-Image-2.6 Preview has no published price card

• Licence — M​OSS-Image is Apache 2.0 for both weights and standalone pipeline code, with no revenue threshold; M​AI-Image-2.6 Preview is proprietary with no stated licence

• Independent score — M​OSS-Image second on RealLayoutArena at 1,698.70 Elo (author-built benchmark); M​AI-Image-2.6 Preview second on Arena text-to-image at 1,336±11 on ~3,488 votes (crowd board), third on Arena single-image edit at 1,420

• Layout control — M​OSS-Image accepts a JSON layout with absolute pixel coordinates and reports 0.88 mean IoU on 7Bench; M​icrosoft publishes no layout benchmark for anything in the M​AI image line

• Text rendering — M​OSS-Image reports 96.3 Mean-6 on LongLongTextBench with 97.1 English and 92.4 Chinese; M​AI-Image-2.6 Preview gained +91 Elo on text rendering over 2.5 per Arena's category breakdown, with no absolute figure published

• Editing — M​AI-Image-2.6 Preview is measured at No. 3 on Arena's single-image-edit board; M​OSS-Image's report does not present editing as a product surface

• Serving — M​OSS-Image has Day 0 framework work in progress, including an open vLLM-Omni pull request that adds the model with skip-layer guidance and is not yet merged into a tagged release; M​AI-Image-2.6 Preview is served by M​icrosoft, on M​icrosoft's terms, when M​icrosoft says so

Where each one is genuinely ahead

M​OSS-Image is ahead on everything that follows from being open. You can read the architecture, quantise it, fine-tune it on private data, run it air-gapped, and ship it inside a commercial product without a licence negotiation — Apache 2.0 covers the weights and the standalone pipeline code, with no acceptable-use annex and no separate commercial tier. For a model whose headline capability is rendering legible text in specific positions, that matters more than a benchmark delta, because text rendering is exactly the capability that shows up in client work: product mockups, interface comps, infographics, posters, ad layouts.

M​AI-Image-2.6 Preview is ahead on the things that follow from being a platform model rather than a repository. It has distribution to hundreds of millions of people through Bing Image Creator, PowerPoint and OneDrive, it is aimed at a managed cloud endpoint on Foundry, and it is being improved on a published cadence. If your requirement is an image model your organisation can call from a M​icrosoft-shaped stack with a support contract behind it, no amount of Apache 2.0 replaces that.

The uncomfortable middle is where most teams actually sit. M​icrosoft has not published a price, a licence or a parameter count, so you cannot model the cost of adopting its model. OpenMOSS has published all of that, and no way to call it without owning hardware.

What neither of them changes about your day

Both of these models sit outside the hosted market you probably already build against, and that is where a routing layer is worth saying something precise rather than promotional. We do not host M​OSS-Image and we do not host M​AI-Image-2.6 Preview, so nothing here is an offer to serve either one. What we do serve is the comparison set: the hosted image models you would measure a self-hosted M​OSS-Image against are reachable through one API key on our catalogue at provider list price with 0% markup, which means a vendor price change on the hosted side is live the same day rather than on a contract cycle, with automatic failover across providers so a benchmarking run does not die because one upstream had a bad afternoon. Trying an unproven checkpoint against a rented one should not require a second SDK, and that is the only claim being made.

For M​AI-Image-2.6 Preview specifically there is nothing to route yet. The honest position is that the model is real, the rank is real, and the access is a waiting list.

Who should act, and who should keep watching

Reach for M​OSS-Image if you have GPUs, a workflow that can express intent as boxes and strings, and a use case where the layout is the deliverable rather than a happy accident. Start with the SFT checkpoint, budget for the encoder's 16.8 GB before you budget for the generator, and expect 50 sampling steps at two transformer forwards per step unless you adopt one of the community distillations. Two of those appeared within a day of release: a 4-step DMD LoRA and an 8-step distilled variant, neither vendor-supported and neither benchmarked.

Wait on M​OSS-Image if you need a callable endpoint this week, if you are renting capacity rather than owning it, or if you want one number that was not produced by the people who built the test. All three of the strong results above come from benchmarks the vendor authored, using a specification stage that the report states runs on G​PT-6 at inference time — so the end-to-end Elo is a two-model pipeline, not a single model's output.

Wait on M​AI-Image-2.6 Preview if you need anything contractual. Watch it if you are already a Foundry customer, because the cadence says the next version is six to eight weeks out and the preview will convert into something with a price attached.

What would settle this comparison is the same thing for both, and neither has it: an independent run of the benchmark each one is ranked on. For M​OSS-Image that means someone other than OpenMOSS running LongLongTextBench or RealLayoutArena. For M​AI-Image-2.6 Preview it means the Arena board accumulating enough votes that ±11 stops mattering, or M​icrosoft publishing a price.

Until one of those happens, the accurate summary is this — a permissively licensed 5B research model with an unusually good text-and-layout interface and no endpoint, against a proprietary platform model with an independent crowd rank, real distribution, and no price. The one you can act on depends entirely on whether your bottleneck is a licence or a URL.

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.