LiteRT Model Catalog

The @localmode/litert package ships a curated catalog (LITERT_MODELS) with three models, all verified to load and generate end-to-end with @litert-lm/core@^0.12.1 in real Chrome.

Gemma 4 E2B and Gemma 4 E4B are the two models Google officially lists as supported by the LiteRT-LM JS API — they use the web-optimized *-it-web.litertlm builds published for browser WebGPU loading. Qwen3 0.6B is a small general .litertlm model included as a lightweight option.

Catalog

Gemma 4 E2B

import { litert } from '@localmode/litert';

const model = litert.languageModel('gemma-4-E2B');

• Size: 2.0GB (gemma-4-E2B-it-web.litertlm)
• Context: 8192 tokens
• License: Gemma
• Backend: WebGPU only (GPU-compiled build — cannot run on CPU)
• Status: Verified end-to-end on Chrome 145 with WebGPU (2026-05-20)
• Best for: The default recommendation — one of the two models Google officially supports for the JS API

Gemma 4 E4B

const model = litert.languageModel('gemma-4-E4B');

• Size: 3.0GB (gemma-4-E4B-it-web.litertlm)
• Context: 8192 tokens
• License: Gemma
• Backend: WebGPU only (GPU-compiled build — cannot run on CPU)
• Status: Verified end-to-end on Chrome 145 with WebGPU (2026-05-20)
• Best for: Higher quality than E2B when the larger download is acceptable

Qwen3 0.6B

const model = litert.languageModel('qwen3-0.6B');

• Size: 614MB (Qwen3-0.6B.litertlm)
• Context: 4096 tokens
• Parameters: ~600M
• License: Apache-2.0
• Backend: WebGPU or CPU (portable build — verified on both)
• Status: Verified end-to-end on Chrome 145 (2026-05-20)
• Best for: Fast loading, quick prototyping, and the only catalog model that runs on the CPU backend

Gated models — not in catalog

The following .litertlm files exist on HuggingFace but are not in the curated catalog: they live behind a HuggingFace account + a click-through Gemma license, which a browser-side fetch() cannot complete. To use them, accept the Gemma license once on the model's HuggingFace page, mint a User Access Token, and load via the modelUrl override.

Loading pattern

Resolve the gated file yourself with an Authorization header, then hand the resulting URL to litert.languageModel() via modelUrl:

import { litert } from '@localmode/litert';

// 1. Accept the Gemma license on HuggingFace, then mint an Access Token.
const HF_TOKEN = import.meta.env.VITE_HF_TOKEN;

// 2. Fetch with the token so HuggingFace returns a signed redirect URL.
const response = await fetch(
  'https://huggingface.co/litert-community/Gemma3-1B-IT/resolve/main/Gemma3-1B-IT_multi-prefill-seq_q8_ekv2048.litertlm',
  { headers: { Authorization: `Bearer ${HF_TOKEN}` } },
);

// 3. Pass the resolved URL to LiteRT.
const model = litert.languageModel('gemma-3-1B', { modelUrl: response.url });

Programmatic Access

Access the catalog at runtime via LITERT_MODELS:

import { LITERT_MODELS, getModelCategory } from '@localmode/litert';
import type { LiteRTModelId } from '@localmode/litert';

for (const [id, info] of Object.entries(LITERT_MODELS)) {
  const category = getModelCategory(info.sizeBytes);
  console.log(`[${category}] ${info.name}: ${info.size}`);
}

const modelId: LiteRTModelId = 'gemma-4-E2B';
const entry = LITERT_MODELS[modelId];
console.log(entry.url); // HuggingFace .litertlm URL

The LiteRTModelEntry shape:

interface LiteRTModelEntry {
  name: string;
  contextLength: number;
  sizeBytes: number;
  size: string;           // e.g. '2.0GB'
  description: string;
  url: string;            // HuggingFace .litertlm URL
  parameterCount: number;
  requiresWebGPU?: boolean; // true = GPU-compiled build, cannot run on CPU backend
}

For per-instance overrides used when loading non-catalog models, see LiteRTModelSettings — notably modelUrl and contextLength — documented on the Overview page.

Next Steps