WebAI

Side by side

Compare runtimes

Send one prompt to two or three of your installed models at once and watch the replies and timings side by side.

Setup

Comparison setup

Structured output
What to compare

Up to 3 run at once, which can hold several models in memory at the same time. wllama uses WebGPU when it can and reports which backend each lane actually used.

Pick at least two. Install more from the Models page if you need them.

Adapter evidence

Structured output and speculative decoding

AdapterStructured outputMTP / speculative decoding
wllamawllama 3.5.1 / llama.cpp b9640-dd4623ajson-schema-constrained, grammar-constrainedThe pinned wllama wrapper forwards JSON Schema response_format and raw GBNF grammar constraints to llama.cpp.unsupportedReason: companion-mount-not-exposedThe wrapper cannot mount a separate MTP head as distinct from target GGUF shards.
Transformers.jsTransformers.js 4.2.0 / ONNX Runtime Web 1.26.0-dev.20260416-b7804b056cprompt-and-validateVersion 4.2.0 has no JSON Schema or grammar decoder constraint. WebAI can prompt for JSON and validate the returned value.unsupportedReason: generation-surface-absentThe pinned JavaScript generation surface has no assistant-model, draft-model, speculative-decoding, or MTP option.
LiteRT-LMLiteRT-LM 0.14.0 early previewprompt-and-validateThe preview exposes no arbitrary JSON Schema or grammar argument through this adapter. Schema tests use prompt-and-validate.unsupportedReason: javascript-toggle-absentThe native engine supports Gemma 4 MTP, but 0.14.0's JavaScript AdvancedSettings binding does not expose the native speculative-decoding toggle.
Chrome Prompt APIBrowser-managed Gemini Nanojson-schema-constrainedChrome accepts a JSON Schema responseConstraint on each Prompt API request. GBNF is not exposed.unknownReason: browser-control-absentChrome does not expose whether Gemini Nano uses speculative decoding internally, and WebAI cannot toggle or measure it.

WebAI only calls something supported if the app can turn it on and the difference shows up in a repeatable A/B run. Whatever the browser does internally doesn't count.