Browser Support

Dendrite auto-selects the fastest inference path available: WebGPU (WebLLM) where supported, Transformers.js WASM as a fallback, and an optional cloud API proxy.

Supported Browsers

Browser	Version	WebGPU (fast)	WASM fallback
Chrome	113+	Yes	Yes
Edge	113+	Yes	Yes
Brave	113+	Yes	Yes
Safari	18+	Requires enabling in Feature Flags	Yes
Firefox	—	No	Yes (small models)
Mobile Chrome	—	No (auto-falls back)	Requires apiUrl in 'auto' mode (local inference disabled on mobile)
Mobile Safari	—	No	Requires apiUrl in 'auto' mode

On mobile in 'auto' mode, local inference is disabled (memory pressure crashes browsers) — Dendrite throws MOBILE_UNSUPPORTED unless apiUrl + username + instanceSlug are configured, in which case it uses the cloud proxy. You can force inference: 'fallback' to attempt Transformers.js WASM with a small model anyway.

Checking Support

import { checkWebGPU } from '@agent-layer-zero/dendrite'

const { available, reason } = checkWebGPU()

if (!available) {
  console.log(reason)
  // "Firefox does not support WebGPU. Try Chrome or Edge."
  // "Enable WebGPU in Safari: Settings > Feature Flags > WebGPU"
  // "Mobile browsers do not support WebGPU. Try Chrome on desktop."
}

Storage Requirements

Model weights are downloaded once and cached in IndexedDB:

Model	Download Size	Cache Size
SmolLM2 360M	~250 MB	~350 MB
Qwen3 0.6B	~400 MB	~500 MB
Llama 3.2 1B	~700 MB	~900 MB
Qwen3 1.7B	~1.1 GB	~1.5 GB
Gemma 2 2B (default)	~1.3 GB	~1.7 GB
Llama 3.2 3B	~1.8 GB	~2.3 GB
Qwen3 4B	~2.5 GB	~3.2 GB
Llama 3.1 8B	~4 GB	~5.0 GB
Qwen3 8B	~4.4 GB	~5.5 GB
Gemma 2 9B	~5.2 GB	~6.5 GB

Cache Recovery

If the model cache gets corrupted (e.g., interrupted download), use:

import { deleteAllModelCaches } from '@agent-layer-zero/dendrite'

await deleteAllModelCaches()
// Then reload the page or call neuron.setModel() again

Dendrite also auto-recovers on Cache.add() errors — it clears the cache and reloads the page once automatically.

Error Handling

import { classifyError } from '@agent-layer-zero/dendrite'

try {
  // ... model loading or generation
} catch (err) {
  const { type, message, canClearCache } = classifyError(err)

  switch (type) {
    case 'quota':
      // Storage full — offer to clear cache or pick smaller model
      break
    case 'network':
      // Download failed — check connection
      break
    case 'webgpu':
      // Browser doesn't support WebGPU
      break
    case 'gpu_pipeline':
      // GPU can't run the compute shaders
      break
    case 'unknown':
      // Something else
      break
  }

  if (canClearCache) {
    await deleteAllModelCaches()
  }
}