SmolLM2 360M
- Download
- ~220 MB
- Estimated GPU / system memory
- 400 MB / 700 MB
- Device support
- Requires WebGPU
Local model guide · no account
Choose a local LLM for your device
Compare small starting models, their downloads, and the memory they need. Use WebGPU when your browser supports it, or choose LFM2.5 for a slower CPU fallback. The chat is free and needs no account.
Three starting points
Download and memory estimates come from the same model information as the chat selector. Actual usage depends on your browser, context, and other running apps.
LFM2.5 350M
- Download
- ~255 MB GPU / ~1.5 GB CPU
- Estimated GPU / system memory
- 450 MB / 4.0 GB
- Device support
- WebGPU or slower CPU fallback
Qwen 3.5 0.8B
- Download
- 430 MB
- Estimated GPU / system memory
- ~1.6 GB / ~2.0 GB
- Device support
- Requires WebGPU
Make a practical choice
Balance download size and device memory
-
Start with a small download
Try a compact model for short questions before downloading a larger Qwen model. The selector shows each model's download size.
-
Check device compatibility
WebLLM models require WebGPU. The selector checks your device and marks models that may be slow or unsupported.
-
CPU fallback when needed
LFM2.5 can run through WebAssembly when WebGPU is unavailable. Its CPU version needs a larger download (about 1.5 GB) and more working memory, and can be substantially slower.
-
Leave memory for the conversation
Downloaded weights also need working memory for inference and context. Memory estimates are approximate; close other heavy tabs if loading fails.
Three steps
From model choice to your first answer
-
Step 1
Check the model selector
Open Chat and compare download sizes and device compatibility. Selecting an uncached model asks for your download approval.
-
Step 2
Approve setup
Choose a model, write your first prompt, or use Load Model to begin setup. Approve the download and keep the tab open while progress is shown.
-
Step 3
Try a short question
Check the answer and speed on your device. Cached files can be reused; offline use requires all model and runtime assets to be cached.
For the chat's document, voice, and privacy features, see Free AI chat with no sign-up. Comparing cloud assistants? See WKO AI vs ChatGPT.
Questions
Frequently asked questions
-
Which local chat model should I try first?
Start with SmolLM2 360M for a small WebGPU model, or LFM2.5 if you need a CPU fallback. Use the model selector's device compatibility labels before trying a larger Qwen model.
-
Can I run a model without WebGPU?
LFM2.5 supports a WebAssembly CPU fallback with an approximately 1.5 GB download and about 4 GB working-memory estimate. WebLLM models, including SmolLM2 and Qwen, require WebGPU. CPU speed depends on your device and can be substantially slower.
-
Is download size the same as required memory?
No. Inference needs memory for model weights, intermediate calculations, and the conversation context. The estimates on this page are guidance, not a guarantee that a model will fit on every device.
-
Does switching models require an account or upload my chat?
No account is needed. Core inference stays in your browser when you switch models. An uncached model requires your download approval. Only an external search provider you enable receives its search request.
Try a model on your device
Check compatibility in the selector, approve the model download, and try a short question. No account is required.