Reference

Presets Reference

The @inbrowser/model/local export: the bundled ModelPreset values and the static types they carry.

This page describes the six bundled ModelPreset values exported from @inbrowser/model and the static types they carry.

Spread a preset into createEngine to construct a running engine; see ./engine.md.

import { createEngine, smollm2_360m } from '@inbrowser/model/local';

const engine = createEngine(smollm2_360m);

Bundled presets

Every bundled preset uses dtype: 'q4f16' and backend: 'auto'.

Preset	`model.modelId`	`contextWindow`	`supportsTools`	`supportsThinking`	`supportsVision`	`supportsAudio`	`thinkingTags`
`gemma4_E2B`	`onnx-community/gemma-4-E2B-it-ONNX`	`128000`	`false`	`true`	`false`	`true`	none
`gemma4_E4B`	`onnx-community/gemma-4-E4B-it-ONNX`	`128000`	`false`	`true`	`false`	`true`	none
`smollm2_360m`	`HuggingFaceTB/SmolLM2-360M-Instruct`	`8192`	`false`	`false`	`false`	`false`	none
`qwen2_5_coder_1_5b`	`onnx-community/Qwen2.5-Coder-1.5B-Instruct`	`32768`	`true`	`false`	`false`	`false`	none
`qwen3_1_7b`	`onnx-community/Qwen3-1.7B-ONNX`	`32768`	`true`	`false`	`false`	`false`	none
`deepseek_r1_qwen_1_5b`	`onnx-community/DeepSeek-R1-Distill-Qwen-1.5B-ONNX`	`131072`	`false`	`true`	`false`	`false`	`<think>` / `</think>`

Notes on capability flags:

The Gemma 4 presets declare supportsThinking: true but deliberately omit thinkingTags. Their channel-marker emission is inconsistent, so a text-based parser cannot reliably separate thinking from answer. Reasoning streams inline as token events.
qwen2_5_coder_1_5b and qwen3_1_7b declare supportsTools: true. The engine threads tools into the chat template and wraps the output with parseToolCalls() when tools are provided.
deepseek_r1_qwen_1_5b declares thinkingTags: { openTag: '<think>', closeTag: '</think>' }. These literal-text tags match splitThinking’s defaults; declaring them on the preset lets consumers stay model-agnostic.

`ModelPreset`

A fully-specified model configuration. Spread into createEngine alongside optional EngineHooks.

Field	Type	Description
`model`	`ModelRef`	HF Hub locator.
`dtype`	`Dtype`	Weight/activation precision.
`backend`	`Backend`	ONNX Runtime Web execution backend.
`capabilities`	`EngineCapabilities`	Static, pre-load capability declaration.
`chatTemplate?`	`(messages: ReadonlyArray<EngineMessage>) => string`	Optional chat-template override.

`ModelRef`

Field	Type	Description
`modelId`	`string`	HF Hub repo id.
`revision?`	`string`	Optional revision. The bundled presets pin no revision.

`Dtype`

type Dtype = 'q4f16' | 'q8' | 'fp16' | 'fp32';

Value	Meaning
`q4f16`	4-bit int weights, fp16 activations.
`q8`	8-bit int weights.
`fp16`	Half precision throughout.
`fp32`	Full precision.

`Backend`

type Backend = 'auto' | 'webgpu' | 'wasm';

Value	Meaning
`auto`	Probe `navigator.gpu`; fall back to wasm if absent.
`webgpu`	WebGPU compute pipeline.
`wasm`	SIMD CPU fallback. Always available, much slower.

`EngineCapabilities`

EngineCapabilities is the static capability declaration carried on ModelPreset.capabilities. Its full field table, including thinkingTags, is documented in ./engine.md.

Authoring presets

Community presets are authored with definePreset, the same helper used for the bundled set. See ./engine.md.

Engine surface and shared types: ./engine.md
Adapters and worker: ./adapters-and-worker.md
Tutorial: ../tutorials/01-run-a-model-in-the-browser.md

Bundled presets

ModelPreset

ModelRef

Dtype

Backend

EngineCapabilities