Machine learning in the browser, on WebGPU,
straight from Hugging Face.
Kiln (kiln-webgpu) runs language, vision and speech models on the user's GPU, loading the original safetensors and GGUF files: no ONNX
conversion, no server. It speaks transformers.js's API, and adds output that always matches your schema, LoRA adapters you switch in milliseconds,
retrieval, and one LLM for many tasks.
npm install kiln-webgpu(SmolLM2-135M, q4, same GPU)
names and output shapes
no WASM runtime to download
Quick start
import { load } from "kiln-webgpu";
const llm = await load("Qwen/Qwen2.5-0.5B-Instruct");
for await (const text of llm.chat([{ role: "user", content: "Write a haiku about kilns." }])) {
document.body.append(text);
}The same model as a transformers.js pipeline:
import { pipeline } from "kiln-webgpu";
const generator = await pipeline("text-generation", "Qwen/Qwen2.5-0.5B-Instruct");
const [{ generated_text }] = await generator([{ role: "user", content: "Hello!" }], { max_new_tokens: 64 });Why Kiln
No conversion step
Loads the repo as published: safetensors, pytorch_model.bin or a GGUF file. New fine-tunes and llama.cpp quantizations work the day they appear.
Output that always parses
JSON Schema, regex and grammars are enforced token by token, for text and image prompts alike: read an invoice into your shape, every time.
Conversations get cheaper
Earlier turns are reused across calls and page reloads; draft tokens speed decoding up. A follow-up prefills only what is new.
LoRA adapters in milliseconds
Load a base model once and switch PEFT adapters per call with a GPU copy.
Retrieval in one library
Chunking, embeddings, search and grounded answers with cited passages; the passages' prompt prefix is cached.
One model, many tasks
Classification with your own labels (real probabilities), extraction, NER and QA from a single LLM download.
Checked, not claimed
Outputs are compared with Python transformers and transformers.js on the same inputs; the numbers are published.
Small
One ES module with its worker inside, from a CDN or any bundler. WebGPU where available, a multithreaded CPU fallback where not.
What it runs
| Area | Models | Tasks |
|---|---|---|
| Language | Llama, Mistral, Qwen2/3, SmolLM, Granite 3 and 4, Gemma 3; safetensors or GGUF | chat, completion, tools, structured output, LoRA, text-generation |
| Text encoders | BERT, DistilBERT, RoBERTa, MiniLM, BGE | embeddings, classification, NER, QA, fill-mask, zero-shot |
| Seq2seq | BART, T5 / Flan-T5, Marian, M2M100 / NLLB | summarization, translation, text2text |
| Speech | Whisper, wav2vec2, HuBERT, SpeechT5, MMS-VITS, Kokoro | speech recognition with timestamps, audio classification, text-to-speech |
| Vision | ViT, DeiT, DINOv2, CLIP, ResNet, ConvNeXt, DETR, YOLOS, OWL-ViT / OWLv2, SegFormer, Depth Anything, RMBG, Swin2SR | classification, detection, zero-shot detection, segmentation, depth, background removal, upscaling |
| Multimodal | Gemma 3, SmolVLM, Florence-2, Donut, ViT-GPT2 | chat about images, captions, OCR, document QA |
Requires a browser with WebGPU (Chrome, Edge, Safari 26 and later, Firefox 141 and later on Windows) for full speed; elsewhere Kiln runs on the CPU.