The most-adopted AI design agent in the Figma plugin market does not have a model. Lovart publishes no parameter counts, no training-data volumes, no inference-latency benchmarks, and its repository contains no PyTorch training scripts. It sits in the Figma plugin market's Top 3, gets mentioned in Dribbble designer roundups more often than MidJourney V6, and appears under "Used Tools" on Behance project pages at 2.7 times the rate of Adobe Firefly.
Its first-week retention is 68 percent, against an industry average of 41 percent. That is not a model metric. It is an outcome metric.
If you have spent the last two years assuming the moat is the weights, this reads as an anomaly. It is not. It is the design-domain instance of a shift that is now visible across the whole stack, and the two stories being told about it contradict each other in a way worth sitting with.
The two stories being told about it contradict each other in a way worth sitting with.
The first story is the one the labs have every incentive to tell. OpenAI is valued around $300 billion, Anthropic around $61 billion, and those numbers only pencil out if the frontier model is the irreplaceable asset. Tens of billions in training compute is defensible capital expenditure only if the thing you trained is the thing customers cannot route around.
The second story is the one Perplexity told in February 2026, when it shipped Computer: an agent coordinating nineteen different models, priced at $200 a month. Claude Opus handles orchestration and coding, Gemini handles deep research, Nano Banana handles images, Veo handles video, Grok takes fast lightweight queries, ChatGPT handles long-context recall. Aravind Srinivas has said the orchestration is the product and the model is a tool.
He has also said something sharper: that models, as they specialize, "just become tools similar to the file system, CLI tools, connectors, browser, search." A frontier lab cannot comfortably say this out loud. A lab that admitted the model is a commodity would be undercutting its own valuation narrative, so it routes to its own models by default even when a competitor's model is better for a subtask. Perplexity owns no frontier model it must defend, so it can be indifferent to which lab wins any individual call. That indifference is the product.
Paul Kedrosky put the layer diagram plainly: large language models are the new silicon, orchestration layers like Claude Code and Codex are the new operating systems, applications are the new productivity suite. Model capability progress has slowed, he argues, and improved coordination is masking it. Value migrates up the stack.
For design, that migration has a specific shape, and Lovart is the clearest specimen. Its architecture strings together twenty-seven open-source multimodal models and six commercial APIs through a three-layer middleware: dynamic routing, semantic normalization, output validation. Routing decisions are made by a small XGBoost model with twelve features, deployed on Cloudflare Workers, cold-starting in under eighty milliseconds.
Run the same prompt through Stable Diffusion WebUI and through Lovart and you get two different kinds of output. The WebUI is faster, and its SVG export needs manual tracing. Lovart is about two seconds slower, and what comes out is an editable vector group. Those two seconds buy a deliverable, not a picture.
The more interesting number is 83. In Figma plugin version 2.4.1, 83 percent of requests never trigger image generation at all. They are intercepted at the task-parsing layer and decomposed before any pixels are considered.
Take a request like a 1920×600 homepage banner with a Hanfu cat, a purple-to-pink gradient, "Spring New Arrivals" text in Source Han Sans Bold. The system runs four passes. It identifies this as a page-level requirement rather than a single element, extracts hard constraints like exact dimensions and font, flags soft constraints like whether the Hanfu needs cultural-symbol validation. It atomizes into five parallel subtasks. It schedules each subtask to a different engine. Then it fuses the output through a consistency check: force everything into sRGB, compute gray-distribution entropy to verify the visual weight ratio of cat to text lands in the golden-section range, sample a hundred pixels along the gradient to confirm ΔL* stays under three.
None of that is reasoning in the model sense. It is two thousand hand-annotated design rules, some of which were derived by reverse-parsing 327 award-winning Behance banners through the Figma plugin API. Design common sense, compiled to if-else.
There is also a translator layer that maps fuzzy adjectives into parameters. "High-end feel" becomes saturation at or below 25 percent, contrast ratio at or above 4.5:1, whitespace at or above 35 percent. That mapping was trained on screen recordings of twelve senior UI designers, with an LSTM modeling their operation sequences. The output is not text. It is a Figma API call chain.
Pixel