Workbench

Chat

Talk to a local model and see what it costs: every reply reports the timings, tokens, cache, and context its runtime actually exposes.

Session

Choose a model

Pick a model below, then pick an engine that can run it.

Model

Download a model on the Models page, or use a model your browser already manages.

Engine

Engines appear once you pick a model.

CPU threadsContext (K tokens)This model doesn't list a maximum, so WebAI defaults to 2 K. 1 K = 1,024 tokens; values round up to the next whole K.

WebGPU offloadFull (wllama default)PartialCPU only

Chats

Saved chatChat name

Loading saved chats.

System prompt

Generation settings

TemperatureTop PTop KRepeat penaltySeedMax output tokens

Leave the seed blank for random. Leave max output blank to run until the model stops or fills the context. Both are saved with each reply.

Prompt caching

wllama asks llama.cpp to reuse the longest matching prompt prefix and reports the number of cached tokens for each response.

MTP unavailable

wllama 3.5.1 exposes generic draft-model parameters, but loadModel renames every supplied non-mmproj GGUF as a target shard. It has no browser wrapper surface that mounts an MTP companion separately or selects llama.cpp's MTP path.

Conversation

Chat

No session

Load a model to begin

Every reply comes with load time, TTFT, and prefill and decode speed.