No key.
Nothing to leak, rotate or bill.
Why local
The cloud is a dependency. This isn't.
A key, an account, a server, a signal, a bill. Local chat removes all five. The only thing left is the download.
No server.
The model is the backend.
No signal.
Airplane mode is a supported state.
Live demo
Pull the plug. Ask again.
The first reply downloads the model once. Then turn off Wi-Fi and keep chatting. The answer is generated in this tab.
- 01 Send a question.
- 02 Turn off Wi-Fi.
- 03 Ask another.
This is the instruction the demo sends before your question. Edit it and watch the answer change.
Ask something. The first reply downloads the model once, then answers locally.
Idle. No model loaded.
First send downloads the weights once. About 1.5 GB on CPU or 2.0 GB on GPU. They stay cached and later replies run offline.
What you get
Everything handled. Nothing sent.
Plug and play: the engine, the download, the context window and the cache are the package's job. Bring any UI; your messages, markup and styles stay yours.
GPU when it's there. CPU when it isn't.
WebGPU runs the Gemma 4 E2B build. Without an adapter, the same call falls back to the CPU build. You never branch, detect or configure.
Download once. Reply forever.
About 1.5 GB on CPU or 2.0 GB on GPU. It resumes after a reload, stays cached until you clear it, and then every token is generated on the device.
Your prompt never gets a ticket number.
No key to leak and no server to trust. The image, the prompt and the reply stay in the tab, the worker or the Node process. The package writes no history.
Your UI. Our stream.
Text arrives through callbacks, so React, Svelte, a worker or a terminal render the same reply. Cancel is one call.
It can look at things.
Describe a photo, translate the description, build tags or turn a voice note into a task. Vision ships in the CPU build with no second download.
The boring parts are done.
Engine choice, context limits, download resume, cache reuse and timeouts are handled. You pass messages and render text.
First run
The first reply downloads. The rest are unlimited.
One download, one cache, no server round trip. Every reply after the first is local, offline and unlimited.
Install
npm install @degreesign/chat, then
call createChat. No key, no
account, no config.
Download once
The first send fetches the weights and resumes if the connection drops. 1.5 GB on CPU, 2.0 GB on GPU.
Reply offline, unlimited
Every later send runs on the device. Stream tokens, stop any time, keep the model warm. There is no reply quota.
Proof
DoListy
DoListy ships chat with the network off.
A no-signup PWA for photos, voice notes and tasks.
It captions photos, fills tags and answers questions
with no account and no key. The chat is
@degreesign/chat.
-
Describe
Photos are captioned locally with
describeImage. - Answer Chat replies stream with no account and no API key.
- Ship Works in airplane mode, on a plane, in a basement.
The network is off. The reply still arrives.
Models
You don't pick. It does.
With no override, the package picks the GPU build when WebGPU exists and the CPU build when it does not. Both are 2B models: small enough for a device, good enough for captions, tags and short answers.
Qwen3-VL 2B
CPU · visionThe default on a CPU. Vision travels with the model, so image turns work without a second download.
| Build | Reasoning | Images | Weights |
|---|---|---|---|
qwen3_vl_2b |
Yes | Yes | about 1.5 GB |
qwen3_vl_2b_instruct |
No | Yes | Q4_K_M |
qwen3_vl_2b_instruct_q3 |
No | Yes | Q3, 160 MB less |
4K token context on a phone, 8K on a desktop. The projector shares the text thread budget.
Gemma 4 E2B
WebGPUThe default on a GPU. A 2.0 GB build with a large context for text turns.
| Build | Reasoning | Images | Weights |
|---|---|---|---|
gemma4_e2b_limited |
Yes | No | 2.0 GB |
16K token context on a phone, 32K on a desktop. The build ships without vision or audio encoders.
Straight talk.
These are 2B models. They are excellent at captions, tags, summaries, short answers and private drafts. They are not the model you ask to write a novel. The first reply on a phone can take a minute. After that, it is local.
Technical details
Context and budget
- Reply tokens 250 default
- Images and audio 2 / 1
- Prompt characters 10000
- CPU threads one fifth kept free
Timeouts and transport
- First chunk 5 min desktop / 50 phone
- Image prefill 14 min
- Silence 30 s
- Download range 32 MB
Compared
Less to download. Longer conversations. No setup.
Both packages can run the same Gemma 4 model right in the browser. The difference is what each one asks of the device and of you: the download, how much of the conversation it remembers, and the setup needed to ship. Here is the practical comparison.
2.0 GB
2.8 GB
Same base model, 800 MB less before the first token.
32K
4K
16K on a phone. web-llm caps both at 4K.
86.2 KB
6.8 MB
Minified. The web-llm entry is 6.2 MB after minification.
| Question | @degreesign/chat | @mlc-ai/web-llm |
|---|---|---|
| Smallest usable chat build | Gemma 4 E2B, 2.0 GB | 2.8 GB, smaller builds not usable |
| GPU context | 32K desktop, 16K mobile | 4K |
| Engine and model choice | automatic | host builds it |
| Browser entry | 86.2 KB minified | 6.8 MB, 6.2 MB minified |
| Package without wasm | 1.7 MB, wasm in package | 14.4 MB, wasm at load |
The honest verdict.
Choose @degreesign/chat for
plug-and-play chat on mobile PWAs, phones and consumer
web apps. Choose @mlc-ai/web-llm for
manual, desktop-only setups tuned for
shader-f16 GPUs.
Architecture
No black box.
Every function is on one offline HTML page: the entry, the catalog, the prompt budgets, both engines and the stores. Read it before you ship it.
- 01 Public entry and typed API
- 02 Chat service and model catalog
- 03 Prompt budgets and streaming
- 04 wllama and LiteRT engines
- 05 Stores, assets and weights
Opens in a new tab. One self-contained HTML file. Browse the source .
code structure
full map
Public entry
createChat / preloadModel / describeImage
Model runtime
service / catalog / engine choice
Prompt and engines
budgets / streaming / wllama / LiteRT
Tools and assets
descriptors / wasm / weights
Stores and types
storage / caches / enums
Open full architecture
Install
One command to local.
No key, no account, no server. Add the package, call
createChat, render in any UI. Works in
React, a worker, a Node script or plain JS.
npm install @degreesign/chat
No config, no key. The first call downloads the weights once; later calls reuse the cache.
import {
LlmRole,
createChat,
} from "@degreesign/chat";
const cancel = createChat({
messages: [
{
role: LlmRole.system,
content: "Be brief.",
},
{
role: LlmRole.user,
content: "Hello",
},
],
onResponse: (text) => render(text),
onStatus: (state) => show(state),
});
cancel();FAQ
The questions you're about to ask.
No jargon, no asterisks. Here is what it does and what it does not.
Do I need a GPU?
No. WebGPU is used when the browser has it. Without an adapter, the CPU engine runs the same call. Phones and older laptops work.
How big is the download?
About 1.5 GB on CPU or 2.0 GB on GPU. It downloads once, resumes after a reload, and stays cached until you clear it.
Does it really work offline?
Yes. After the first download, the weights and the engine run locally. The package makes no network request to answer.
What happens if I clear the cache?
The next send downloads it again. No hidden copy, no account.
How is it free?
The package is MIT and there is no per-reply fee. You pay for the device and the one-time download, not for each answer.
Does it collect anything?
No key, no server and no upload. The package writes no history; your app owns the data.
Start here
Install once. Answer offline.
One package. Two engines. Zero keys. Unlimited offline replies.
MIT · Types included · 1.5 GB CPU / 2.0 GB GPU