Transformers.js v4.3: Structured Output in the browser

This Hugging Face video walks through the Transformers.js 4.3 release, whose headline feature is structured output in the browser. A new package, @huggingface/transformers-structured-output, forces a model to follow an exact JSON Schema or regular expression.

The first minute and a half covers the rest of the release. After the announcement at 0:00, the chapter at 0:26 is about the new documentation pipeline, which the description calls a complete overhaul. Community-contributed fixes and features get a short mention at 1:00.

From 1:19 the video sets out the problem: language models return unreliable JSON, and developers end up stripping markdown fences and hoping the property names came out right. At 2:18 it explains how constrained decoding works. According to the description, the new package hooks into the logits processor in Transformers.js; adding that processor is the subject of the chapter at 3:25.

The longest stretch, from 3:37 to 6:03, covers the new packages folder in the repository and the plugin approach that structured output uses. At 6:03 the video turns to constraining output with a regex instead of a JSON Schema, and at 6:45 a recipe generator demo shows the feature in use.

The second half tells the story behind the feature. From 7:33 it covers llguidance, which the first version wrapped through WASM. At 9:02 the video explains rewriting it in pure JavaScript, which the description says runs with almost no overhead. A chapter at 9:22 looks at performance on Gemma 4, Granite and LFM 2.5; the sources do not give the figures. The last chapter, at 10:26, points viewers to try it.

The description links the 4.3.0 release notes on GitHub and the npm pages for both the main library and the structured output package.