Models

This table is generated from the registry in the package, so it matches the version you install. Filter by verb or input. Rows you cannot run in this browser are dimmed.

Stable only

Model	Verb	Accepts	Download	Runs on	Status	Licence
lfm2.5-230m 230M params alias `text:tiny` · Smallest LFM2.5 language model. Rewrites, extraction, short answers.	`generate`	text	201 MB	WebGPU CPU WASM	stable	`lfm1.0`
lfm2.5-350m 350M params alias `text:default` · Fast general model with native tool calling.	`generate`	text	243 MB	WebGPU CPU WASM	stable	`lfm1.0`
lfm2.5-1.2b 1.2B params General chat and reasoning. Needs WebGPU in browsers.	`generate`	text	725 MB	WebGPU CPU	preview	`lfm1.0`
qwen3-0.6b 0.6B params Multilingual model with an optional reasoning mode.	`generate`	text	544 MB	WebGPU WASM CPU	preview	`apache-2.0`
gemma-3-270m 270M params Tiny Gemma. Uses fp32 on WebGPU because fp16 overflows on that backend.	`generate`	text	308 MB	WebGPU CPU WASM	preview	`gemma`
functiongemma-270m 270M params Gemma trained for function calling. Best after fine-tuning on your own tools.	`generate`	text	406 MB	WebGPU WASM CPU	preview	`gemma`
lfm2-1.2b-tool 1.2B params LFM2 tuned for tool use.	`generate`	text	828 MB	WebGPU WASM CPU	preview	`lfm1.0`
lfm2.5-vl-450m 450M params alias `vision:default` · Small vision-language model. Questions about images, screenshots and short clips.	`generate`	text image video	644 MB	WebGPU CPU WASM	stable	`lfm1.0`
lfm2.5-vl-1.6b 1.6B params Larger vision-language model for better answers on GPU devices.	`generate`	text image video	1.6 GB	WebGPU CPU	preview	`lfm1.0`
smolvlm-256m 256M params Tiny vision-language model.	`generate`	text image	253 MB	WebGPU WASM CPU	preview	`apache-2.0`
florence-2-base 230M params alias `vision:fast` · Captions, OCR and object detection through presets.	`generate`	image text	725 MB	WebGPU WASM CPU	stable	`mit`
moonshine-tiny 27M params alias `stt:tiny` · Tiny English speech recognizer for live captions on any device.	`generate`	audio	27 MB	WebGPU WASM CPU	stable	`mit`
moonshine-base 61M params alias `stt:realtime` · English speech recognizer with better accuracy than Moonshine tiny.	`generate`	audio	60 MB	WebGPU WASM CPU	stable	`mit`
whisper-tiny-en 39M params Whisper tiny, English only, with timestamps.	`generate`	audio	39 MB	WebGPU WASM CPU	preview	`mit`
whisper-base 74M params alias `stt:multilingual` · Small multilingual transcription with timestamps.	`generate`	audio	70 MB	WebGPU WASM CPU	preview	`mit`
whisper-small 244M params Multilingual transcription with timestamps, more accurate than Whisper base.	`generate`	audio	226 MB	WebGPU WASM CPU	preview	`mit`
whisper-large-v3-turbo 809M params alias `stt:accurate` · Multilingual transcription with timestamps.	`generate`	audio	724 MB	WebGPU WASM CPU	preview	`mit`
chrome:gemini-nano Chrome's built-in model through the Prompt API. No download for your app.	`generate`	text	managed by Chrome	WebGPU WASM	preview	`google-terms`
nli-deberta-v3-xsmall 22M params alias `judge:router` · Zero-shot judge: scores text against option descriptions with natural-language inference.	`evaluate`	text	83 MB	WebGPU WASM CPU	stable	`apache-2.0`
prompt-injection-deberta-v3 184M params alias `judge:injection` · Detects prompt injection and jailbreak attempts.	`evaluate`	text	705 MB	WebGPU WASM CPU	stable	`apache-2.0`
bert-base-ner 110M params alias `judge:entities` · Named entities: people, organisations, locations. English.	`evaluate`	text	104 MB	WebGPU WASM CPU	stable	`mit`
piiranha-v1 280M params PII detection in 6 languages. Non-commercial licence.	`evaluate`	text	302 MB	WebGPU WASM CPU	preview	`cc-by-nc-nd-4.0`
lfm2.5-encoder-350m-router 350M params LFM2.5 zero-shot router: scores text against any lanes in one pass. 15 languages, 8K context.	`evaluate`	text	429 MB	WebGPU WASM CPU	stable	`lfm1.0`
lfm2.5-encoder-350m-pii 350M params alias `judge:pii` · LFM2.5 PII detector: 40 entity types (contact, credential, financial, health and more) in 16 languages.	`evaluate`	text	565 MB	WebGPU WASM CPU	stable	`lfm1.0`
d1-omni-600m 587M params alias `judge:omni` · LiquidAI d1 decision model: typed answers over text, images or speech in one pass, zero output tokens.	`evaluate`	text image audio	856 MB	WebGPU WASM CPU	stable	`lfm1.0`
d1-3b 3.1B params alias `judge:large` · LiquidAI d1 decision model, 3B: the most accurate d1, over text and images, answered at the last token.	`evaluate`	text image	4.3 GB	WebGPU CPU	preview	`lfm1.0`
all-minilm-l6-v2 22M params alias `embed:tiny` · Tiny English embedding model. 384 dimensions.	`embed`	text	22 MB	WebGPU WASM CPU	stable	`apache-2.0`
embeddinggemma-300m 300M params alias `embed:default` · Multilingual embeddings. 768 dimensions, shortenable to 128.	`embed`	text	188 MB	WebGPU CPU WASM	stable	`gemma`
kokoro-82m 82M params alias `voice:default` · High-quality English speech with 28 voices.	`speak`	text	88 MB	WebGPU WASM CPU	stable	`apache-2.0`
chatterbox-turbo 350M params alias `voice:clone` · Zero-shot voice cloning from 3 to 10 seconds of reference audio. English. Needs WebGPU in browsers.	`speak`	text audio	684 MB	WebGPU CPU	preview	`mit`
sd-turbo 1B params alias `image:default` · One-step image generation at 512×512.	`paint`	text	2.4 GB	WebGPU CPU	preview	`stabilityai-community`
chronos-bolt-tiny 9M params alias `forecast:default` · Zero-shot time-series forecaster (Amazon Chronos-Bolt tiny). 64 steps per pass, 9 quantiles.	`forecast`	series	33 MB	WebGPU WASM CPU	stable	`apache-2.0`
chronos-bolt-small 48M params Larger Chronos-Bolt with better accuracy on complex series.	`forecast`	series	182 MB	WebGPU WASM CPU	stable	`apache-2.0`
silero-vad 2M params alias `vad:default` · Voice activity detection used by mic() and live transcription.	`vad`	audio	2 MB	WASM CPU	stable	`mit`

Status#

stable: real-model tests pass on Bun, Node and Chromium. Available by default.
preview: works, but is tested on fewer runtimes or has known limits. Needs allowPreview: true.
experimental: new or unusual hosting. Needs allowPreview: true, and may change.

Aliases#

Aliases pick a good default for a job. They may point to a better model in a minor release; use the full ID to pin one.

Alias	Resolves to
`text:tiny`	`lfm2.5-230m`
`text:default`	`lfm2.5-350m`
`vision:default`	`lfm2.5-vl-450m`
`vision:fast`	`florence-2-base`
`stt:realtime`	`moonshine-base`
`stt:tiny`	`moonshine-tiny`
`stt:accurate`	`whisper-large-v3-turbo`
`stt:multilingual`	`whisper-base`
`judge:router`	`nli-deberta-v3-xsmall`
`judge:injection`	`prompt-injection-deberta-v3`
`judge:entities`	`bert-base-ner`
`judge:pii`	`lfm2.5-encoder-350m-pii`
`judge:omni`	`d1-omni-600m`
`judge:large`	`d1-3b`
`embed:default`	`embeddinggemma-300m`
`embed:tiny`	`all-minilm-l6-v2`
`voice:default`	`kokoro-82m`
`voice:clone`	`chatterbox-turbo`
`image:default`	`sd-turbo`
`vad:default`	`silero-vad`
`forecast:default`	`chronos-bolt-tiny`

Where models come from#

Most models are ONNX exports on Hugging Face, pinned to a commit.
Models with no official ONNX export (Chronos-Bolt, the LiquidAI LFM2.5 encoders and LiquidAI d1-omni) were exported by Edgewise and are served from the kucukkanat/edgewise-models Hugging Face bucket. Each file is checked against a SHA-256 hash on download.

The export scripts are in scripts/export in the repository.

Licences#

Each model keeps its own licence, shown in the table. Restrict what your app can load:

configure({ licenses: ['apache-2.0', 'mit'] });