Examples

Each runs entirely in your browser. The first run downloads the model from the Hugging Face Hub and caches it; later visits load from the cache. A GPU with WebGPU is strongly recommended.

Streaming chat

A chat that streams tokens as they decode. Each turn reuses the cached prompt of the last one, so only the new message is prefilled.

Qwen2.5-0.5B-Instruct, about 500 MB on first use

Text to JSON

Extraction into a schema you write. Kiln's grammar enforces the schema at every token, so the result always parses and fits.

Qwen2.5-0.5B-Instruct

Invoice image to JSON

A vision-language model reads a document image straight into your schema. The shape is guaranteed; the values are the model's reading, so check them.

SmolVLM-500M-Instruct, about 1 GB on first use

Ask your documents

Paste documents (separated by blank lines), ask a question: Kiln embeds the passages, finds the closest, and answers citing them as [1], [2].

all-MiniLM-L6-v2 + Qwen2.5-0.5B-Instruct

One model, many tasks

Classification with labels you make up, and named entities of types you choose, from one LLM. Label scores are the model's own probabilities.

Qwen2.5-0.5B-Instruct

Switching LoRA adapters

One base model with two PEFT adapters: answers from the base model, a science-tuned adapter, and a LinkedIn-style rewriter. Switching between loaded adapters is a GPU copy.

SmolLM2-135M-Instruct + two adapters

Speech to text

Whisper transcribes a recording with word-level timestamps, computed as transformers.js computes them.

whisper-tiny.en

Find anything in a photo

Zero-shot object detection: boxes for whatever you type, no training.

OWL-ViT B/32

The playground has a full chat UI with model picker, and the kernel checks.