On-device LLMs &
AI SDK Provider

01import { useEngine } from "@tryhamster/gerbil/gpu/hooks";
02
03function Chat() {
04  // Native WebGPU engine, cached after first download.
05  // No model arg → a device-aware default. Or pass { model: "..." }.
06  const { complete, completion, isGenerating } = useEngine();
07
08  return (
09    <button onClick={() => complete(userInput)} disabled={isGenerating}>
10      {completion || "Generate"}
11    </button>
12  );
13}

01import { WebGPUEngine } from "@tryhamster/gerbil/gpu";
02
03// Same native engine on Node.js (WebGPU via Dawn)
04const engine = await WebGPUEngine.create({
05  repo: "Qwen/Qwen3.5-0.8B", // thinking-capable, 262k context
06});
07
08const result = await engine.generate("Write a haiku", {
09  maxTokens: 100,
10});
11
12console.log(result.thinking); // reasoning steps
13console.log(result.text);     // final response

01import { useEngine } from "@tryhamster/gerbil/gpu/hooks";
02
03function Vision() {
04  // Qwen3.5's own ViT tower (+~192 MB) — bit-exact vs HF.
05  const { describeImage, completion } = useEngine({ enableVision: true });
06
07  // image-in -> text-out (URL, File, or decoded RGB pixels)
08  return (
09    <button onClick={() => describeImage(photoUrl, "What's in this photo?")}>
10      {completion || "Describe"}
11    </button>
12  );
13}

01import { defineTool } from "@tryhamster/gerbil";
02import { z } from "zod";
03
04const weather = defineTool({
05  name: "get_weather",
06  description: "Get weather for a city",
07  parameters: z.object({
08    city: z.string(),
09  }),
10  execute: async ({ city }) => {
11    return `Weather in ${city}: 72°F, sunny`;
12  },
13});
14
15// LLM can now call this tool during generation

01import { commit, summarize, review } from "@tryhamster/gerbil/skills";
02import { describeImage, captionImage } from "@tryhamster/gerbil/skills";
03
04// Generate commit message from staged changes
05const msg = await commit({ type: "conventional" });
06
07// Summarize any content
08const tldr = await summarize({ content: longDoc });
09
10// Vision skills
11const alt = await captionImage({ image: photoUrl });
12const analysis = await describeImage({ 
13  image: screenshot, 
14  focus: "text" 
15});

01import { useTTS } from "@tryhamster/gerbil/gpu/hooks";
02
03function Speaker() {
04  // Native WebGPU TTS (Kani-TTS-2) — pure WGSL, no ONNX.
05  // Synthesizes + plays 22.05 kHz audio on-device.
06  const { speak, isSynthesizing } = useTTS();
07
08  return (
09    <button onClick={() => speak("Hello, I'm Gerbil!")} disabled={isSynthesizing}>
10      Speak
11    </button>
12  );
13}

01import { useSTT } from "@tryhamster/gerbil/gpu/hooks";
02
03function VoiceInput() {
04  // Native WebGPU ASR (Moonshine) — no ONNX, no log-mel.
05  // Captures the mic and resamples to 16 kHz mono for you.
06  const { startRecording, stopRecording, isRecording, transcript } = useSTT();
07
08  return (
09    <div>
10      <button onClick={() => (isRecording ? stopRecording() : startRecording())}>
11        {isRecording ? "Stop" : "Record"}
12      </button>
13      <p>{transcript}</p>
14    </div>
15  );
16}

$ gerbil "Write a haiku about coding"

🤖 Loading lfm2.5-350m...
✓ Model loaded (2.3s)

Silent keystrokes fall
Bugs emerge from tangled code  
Coffee saves the day

⚡ 47.2 tok/s | 0.8s

$ gerbil speak "Hello world" --voice bf_emma
$ gerbil transcribe audio.wav --timestamps
$ gerbil voice question.wav  # STT → LLM → TTS