Built on WebLLM by MLC AI, which uses WebGPU to run quantized LLMs at near-native speed directly in the browser.

What it does

Loads open-source LLMs (Llama, Phi, Gemma, and more) into the browser via WebGPU
Models are downloaded from Hugging Face on first use and cached locally — subsequent loads are instant
Full conversation context support (multi-turn chat with memory)
Swap models on the fly without losing the conversation history
Works completely offline after the first model download
No API key, no account, no backend — just the player's browser

Requirements

WebGPU-capable browser: Chrome 113+, Edge 113+, Safari 18+ — Firefox may work depending on version and GPU
HTTPS or localhost for serving your game (required for the browser cache API that stores model weights — see note below)
Enough VRAM/RAM for the model: smallest models need ~1 GB, larger ones 4 GB+

HTTP note: If you serve your game over plain HTTP (e.g. a LAN IP during development), model caching is disabled and the model re-downloads on each page load. The extension handles this automatically — no crash, just slower cold starts. Use HTTPS for production.

Installation

Download webllm.json from this page
In GDevelop, open your project → Project Manager → Create or search for new extensions → Import extension
Select the downloaded webllm.json
The extension is ready — no other setup needed

Quick start

1. Load the model at scene start

The WebLLM library loads automatically when your scene starts. Add a one-time event at the beginning of your loading scene:

At the beginning of the scene:   → WebLLM: Load Model ("")

Leave the model ID blank to use the default (Llama-3.2-1B-Instruct-q4f32_1-MLC), or set __WebLLM.ModelId in your scene variables to any supported model ID before calling Load Model.

2. Show loading progress

[WebLLM: Is model loading]   → Set text of LoadingLabel to: WebLLM::GetLoadText()

GetLoadText() returns the live status string from WebLLM, e.g.: Loading model from cache [7/108]: 680MB loaded. 15% completed, 4 secs elapsed.

3. Wait until ready, then enable your UI

[WebLLM: Is model ready] [Trigger once]   → Hide LoadingLabel   → Enable your chat input / buttons

4. Attach the LLM behavior and chat

Add the LLM behavior to any object. Then:

[Send button is pressed]   → MyObject: LLM: Send message to LLM (txtInput.Value, "You are a helpful NPC.")    [MyObject: LLM: On message from LLM]   → Set text of ChatLabel to: MyObject.LLM::getResponse()

Extension reference

Extension-level actions & conditions

Name	Type	Description
`Load Model`	Action	Load (or swap) a model by ID. Leave ID blank to use `__WebLLM.ModelId`.
`Is model ready`	Condition	True when model is fully loaded and ready.
`Is model loading`	Condition	True while model is downloading or initializing.
`GetLoadProgress()`	Expression	Loading progress, 0–100.
`GetStatus()`	String expression	`""` / `loading` / `ready` / `error`
`GetLoadText()`	String expression	Human-readable progress text from WebLLM.
`GetAvailableModels()`	String expression	Comma-separated list of all supported model IDs.

Scene variable `__WebLLM`

Child	Type	Description
`ModelId`	String	Model to load (default: `Llama-3.2-1B-Instruct-q4f32_1-MLC`)
`Status`	String	Current status (`loading` / `ready` / `error`)
`LoadProgress`	Number	0–100
`LoadText`	String	Progress text (same as `GetLoadText()`)

LLM behavior

Attach to any object. Handles sending messages and receiving responses asynchronously.

Name	Type	Description
`Send message to LLM`	Action	Single-turn: send text + optional system prompt. Whole response arrives via `On message from LLM`.
`Send messages to LLM (with context)`	Action	Multi-turn: send a context array. Whole response arrives via `On message from LLM`.
`Send message to LLM (streaming)`	Action	Single-turn streaming: tokens arrive one by one via `On delta received`.
`Send messages to LLM with context (streaming)`	Action	Multi-turn streaming: tokens arrive one by one via `On delta received`.
`Add message to context`	Action	Append a `{role, content}` entry to a context array variable.
`On message from LLM`	Condition	Triggers once when the full response has arrived.
`On delta received from LLM`	Condition	Triggers once per token during streaming. Use `getLastDelta()` inside it.
`On error from LLM`	Condition	Triggers once when an error occurs.
`Is generating`	Condition	True while the LLM is currently generating a response.
`getResponse()`	String expression	The full response text (available after `On message from LLM`).
`getLastDelta()`	String expression	The latest streaming token (use inside `On delta received`).
`getError()`	String expression	The last error message.

Multi-turn conversation example (non-streaming)

[Send button pressed]   
→ MyObject: LLM: Add message to context (txtInput.Value, "user", GPTcontext)   
→ MyObject: LLM: Send messages to LLM with context (GPTcontext, "")  
[MyObject: LLM: On message from LLM]   
→ MyObject: LLM: Add message to context (MyObject.LLM::getResponse(), "assistant", GPTcontext)   
→ Set text of ChatLabel to: MyObject.LLM::getResponse()

Multi-turn conversation example (streaming)

[Send button pressed]   
→ MyObject: LLM: Add message to context (txtInput.Value, "user", GPTcontext)   
→ MyObject: LLM: Send messages to LLM with context (streaming) (GPTcontext, "")   
→ Append to ChatLabel: NewLine() + "AI: "  
[MyObject: LLM: On delta received from LLM]   
→ Append to ChatLabel: MyObject.LLM::getLastDelta()  
[MyObject: LLM: On message from LLM]   
→ MyObject: LLM: Add message to context (MyObject.LLM::getResponse(), "assistant", GPTcontext)   
→ Re-enable send button and input

Popular model IDs

Model ID	Size	Notes
`Llama-3.2-1B-Instruct-q4f32_1-MLC`	~800 MB	Smallest Llama 3.2, fast
`Llama-3.2-3B-Instruct-q4f32_1-MLC`	~2 GB	Good balance
`Llama-3.1-8B-Instruct-q4f32_1-MLC`	~5 GB	High quality, needs good GPU
`Phi-3.5-mini-instruct-q4f16_1-MLC`	~2 GB	Microsoft Phi, very capable for its size
`gemma-2-2b-it-q4f32_1-MLC`	~1.5 GB	Google Gemma 2
`TinyLlama-1.1B-Chat-v1.0-q4f32_1-MLC`	~600 MB	Lightest option

Quantization suffixes: q4f16 = 4-bit weights, 16-bit activations; q4f32 = 4-bit weights, 32-bit activations; q0f16 / q0f32 = unquantized full precision (larger, best quality); -1k = 1K context window variant (uses less memory).

~135M parameters

SmolLM2-135M-Instruct-q0f16-MLC
SmolLM2-135M-Instruct-q0f32-MLC

~360M parameters

SmolLM2-360M-Instruct-q0f16-MLC
SmolLM2-360M-Instruct-q0f32-MLC
SmolLM2-360M-Instruct-q4f16_1-MLC
SmolLM2-360M-Instruct-q4f32_1-MLC

~500M parameters

Qwen2-0.5B-Instruct-q0f16-MLC
Qwen2-0.5B-Instruct-q4f16_1-MLC
Qwen2.5-0.5B-Instruct-q0f16-MLC
Qwen2.5-0.5B-Instruct-q4f16_1-MLC
Qwen2.5-0.5B-Instruct-q4f32_1-MLC
Qwen2.5-Coder-0.5B-Instruct-q0f16-MLC
Qwen2.5-Coder-0.5B-Instruct-q4f16_1-MLC
Qwen2.5-Coder-0.5B-Instruct-q4f32_1-MLC

~600M parameters

Qwen3-0.6B-q0f16-MLC
Qwen3-0.6B-q4f16_1-MLC
Qwen3-0.6B-q4f32_1-MLC

~1.1B parameters

TinyLlama-1.1B-Chat-v0.4-q4f16_1-MLC
TinyLlama-1.1B-Chat-v0.4-q4f32_1-MLC
TinyLlama-1.1B-Chat-v0.4-q4f16_1-MLC-1k
TinyLlama-1.1B-Chat-v0.4-q4f32_1-MLC-1k
TinyLlama-1.1B-Chat-v1.0-q4f16_1-MLC
TinyLlama-1.1B-Chat-v1.0-q4f32_1-MLC
TinyLlama-1.1B-Chat-v1.0-q4f16_1-MLC-1k
TinyLlama-1.1B-Chat-v1.0-q4f32_1-MLC-1k
Llama-3.2-1B-Instruct-q0f16-MLC
Llama-3.2-1B-Instruct-q4f16_1-MLC
Llama-3.2-1B-Instruct-q4f32_1-MLC

~1.5B parameters

phi-1_5-q4f16_1-MLC
phi-1_5-q4f32_1-MLC
phi-1_5-q4f16_1-MLC-1k
phi-1_5-q4f32_1-MLC-1k
Qwen2-1.5B-Instruct-q4f16_1-MLC
Qwen2-1.5B-Instruct-q4f32_1-MLC
Qwen2-Math-1.5B-Instruct-q4f16_1-MLC
Qwen2-Math-1.5B-Instruct-q4f32_1-MLC
Qwen2.5-1.5B-Instruct-q4f16_1-MLC
Qwen2.5-1.5B-Instruct-q4f32_1-MLC
Qwen2.5-Coder-1.5B-Instruct-q4f16_1-MLC
Qwen2.5-Coder-1.5B-Instruct-q4f32_1-MLC
Qwen2.5-Math-1.5B-Instruct-q4f16_1-MLC
Qwen2.5-Math-1.5B-Instruct-q4f32_1-MLC

~1.6B parameters

stablelm-2-zephyr-1_6b-q4f16_1-MLC
stablelm-2-zephyr-1_6b-q4f32_1-MLC
stablelm-2-zephyr-1_6b-q4f16_1-MLC-1k
stablelm-2-zephyr-1_6b-q4f32_1-MLC-1k

~1.7B parameters

SmolLM2-1.7B-Instruct-q4f16_1-MLC
SmolLM2-1.7B-Instruct-q4f32_1-MLC
Qwen3-1.7B-q4f16_1-MLC
Qwen3-1.7B-q4f32_1-MLC

~2B parameters

gemma-2b-it-q4f16_1-MLC
gemma-2b-it-q4f32_1-MLC
gemma-2b-it-q4f16_1-MLC-1k
gemma-2b-it-q4f32_1-MLC-1k
gemma-2-2b-it-q4f16_1-MLC
gemma-2-2b-it-q4f32_1-MLC
gemma-2-2b-it-q4f16_1-MLC-1k
gemma-2-2b-it-q4f32_1-MLC-1k
gemma-2-2b-jpn-it-q4f16_1-MLC
gemma-2-2b-jpn-it-q4f32_1-MLC

~2.7B parameters

phi-2-q4f16_1-MLC
phi-2-q4f32_1-MLC
phi-2-q4f16_1-MLC-1k
phi-2-q4f32_1-MLC-1k

~3B parameters

RedPajama-INCITE-Chat-3B-v1-q4f16_1-MLC
RedPajama-INCITE-Chat-3B-v1-q4f32_1-MLC
RedPajama-INCITE-Chat-3B-v1-q4f16_1-MLC-1k
RedPajama-INCITE-Chat-3B-v1-q4f32_1-MLC-1k
Hermes-3-Llama-3.2-3B-q4f16_1-MLC
Hermes-3-Llama-3.2-3B-q4f32_1-MLC
Llama-3.2-3B-Instruct-q4f16_1-MLC
Llama-3.2-3B-Instruct-q4f32_1-MLC
Ministral-3-3B-Base-2512-q4f16_1-MLC
Ministral-3-3B-Reasoning-2512-q4f16_1-MLC
Ministral-3-3B-Instruct-2512-BF16-q4f16_1-MLC
Qwen2.5-3B-Instruct-q4f16_1-MLC
Qwen2.5-3B-Instruct-q4f32_1-MLC
Qwen2.5-Coder-3B-Instruct-q4f16_1-MLC
Qwen2.5-Coder-3B-Instruct-q4f32_1-MLC

~3.8B parameters

Phi-3-mini-4k-instruct-q4f16_1-MLC
Phi-3-mini-4k-instruct-q4f32_1-MLC
Phi-3-mini-4k-instruct-q4f16_1-MLC-1k
Phi-3-mini-4k-instruct-q4f32_1-MLC-1k
Phi-3.5-mini-instruct-q4f16_1-MLC
Phi-3.5-mini-instruct-q4f32_1-MLC
Phi-3.5-mini-instruct-q4f16_1-MLC-1k
Phi-3.5-mini-instruct-q4f32_1-MLC-1k
Phi-3.5-vision-instruct-q4f16_1-MLC
Phi-3.5-vision-instruct-q4f32_1-MLC

~4B parameters

Qwen3-4B-q4f16_1-MLC
Qwen3-4B-q4f32_1-MLC

~7B parameters

DeepSeek-R1-Distill-Qwen-7B-q4f16_1-MLC
DeepSeek-R1-Distill-Qwen-7B-q4f32_1-MLC
Hermes-2-Pro-Mistral-7B-q4f16_1-MLC
Llama-2-7b-chat-hf-q4f16_1-MLC
Llama-2-7b-chat-hf-q4f32_1-MLC
Llama-2-7b-chat-hf-q4f16_1-MLC-1k
Llama-2-7b-chat-hf-q4f32_1-MLC-1k
Mistral-7B-Instruct-v0.2-q4f16_1-MLC
Mistral-7B-Instruct-v0.3-q4f16_1-MLC
Mistral-7B-Instruct-v0.3-q4f32_1-MLC
NeuralHermes-2.5-Mistral-7B-q4f16_1-MLC
OpenHermes-2.5-Mistral-7B-q4f16_1-MLC
Qwen2-7B-Instruct-q4f16_1-MLC
Qwen2-7B-Instruct-q4f32_1-MLC
Qwen2-Math-7B-Instruct-q4f16_1-MLC
Qwen2-Math-7B-Instruct-q4f32_1-MLC
Qwen2.5-7B-Instruct-q4f16_1-MLC
Qwen2.5-7B-Instruct-q4f32_1-MLC
Qwen2.5-Coder-7B-Instruct-q4f16_1-MLC
Qwen2.5-Coder-7B-Instruct-q4f32_1-MLC
WizardMath-7B-V1.1-q4f16_1-MLC

~8B parameters

DeepSeek-R1-Distill-Llama-8B-q4f16_1-MLC
DeepSeek-R1-Distill-Llama-8B-q4f32_1-MLC
Hermes-2-Pro-Llama-3-8B-q4f16_1-MLC
Hermes-2-Pro-Llama-3-8B-q4f32_1-MLC
Hermes-2-Theta-Llama-3-8B-q4f16_1-MLC
Hermes-2-Theta-Llama-3-8B-q4f32_1-MLC
Hermes-3-Llama-3.1-8B-q4f16_1-MLC
Hermes-3-Llama-3.1-8B-q4f32_1-MLC
Llama-3-8B-Instruct-q4f16_1-MLC
Llama-3-8B-Instruct-q4f32_1-MLC
Llama-3-8B-Instruct-q4f16_1-MLC-1k
Llama-3-8B-Instruct-q4f32_1-MLC-1k
Llama-3.1-8B-Instruct-q4f16_1-MLC
Llama-3.1-8B-Instruct-q4f32_1-MLC
Llama-3.1-8B-Instruct-q4f16_1-MLC-1k
Llama-3.1-8B-Instruct-q4f32_1-MLC-1k
Qwen3-8B-q4f16_1-MLC
Qwen3-8B-q4f32_1-MLC

~9B parameters

gemma-2-9b-it-q4f16_1-MLC
gemma-2-9b-it-q4f32_1-MLC

~13B parameters

Llama-2-13b-chat-hf-q4f16_1-MLC

~70B parameters (needs high-end GPU / lots of RAM)

Llama-3-70B-Instruct-q3f16_1-MLC
Llama-3.1-70B-Instruct-q3f16_1-MLC

Embedding models (for semantic/vector search, not chat)

snowflake-arctic-embed-s-q0f32-MLC-b4
snowflake-arctic-embed-s-q0f32-MLC-b32
snowflake-arctic-embed-m-q0f32-MLC-b4
snowflake-arctic-embed-m-q0f32-MLC-b32

To let players choose a model at runtime, use WebLLM::GetAvailableModels() which returns all supported IDs as a comma-separated string, then call Load Model with the chosen ID — the existing conversation context is preserved.

Demo project

A complete working demo scene is available as a paid download ($5). It includes:

Full chat UI wired up to the LLM behavior
Live loading progress bar with status text
System prompt input so you can give the AI a persona
Model switcher — change models mid-conversation without losing context
Properly commented GDevelop events showing every feature of the extension

The demo is a ready-to-open GDevelop folder project. Great as a starting point or just to see how everything fits together.

License

The extension (webllm.json) is free to use in any project, commercial or otherwise.

The underlying WebLLM library is MIT licensed. Individual model weights are subject to their own licenses (Llama models require accepting Meta's license on Hugging Face, Gemma models require Google's license, etc.).

More information

Published	6 hours ago
Status	Released
Category	Tool
Platforms	HTML5
Author	Avram
Made with	GDevelop
Tags	ai, gpt, llm, local, offline, webgpu
AI Disclosure	AI Assisted, Code

Download

Download NowName your own price

Click download now to get access to the following files:

webllm.json 44 kB

WebLLM-example-project.zip 63 kB

if you pay $5 USD or more

WebLLM GDevelop Extension

What it does

Requirements

Installation

Quick start

Extension reference

Extension-level actions & conditions

Scene variable `__WebLLM`

LLM behavior

Multi-turn conversation example (non-streaming)

Multi-turn conversation example (streaming)

Popular model IDs

Demo project

License

Download

Leave a comment

WebLLM GDevelop Extension

What it does

Requirements

Installation

Quick start

Extension reference

Extension-level actions & conditions

Scene variable __WebLLM

LLM behavior

Multi-turn conversation example (non-streaming)

Multi-turn conversation example (streaming)

Popular model IDs

Demo project

License

Download

Leave a comment

Scene variable `__WebLLM`