Skip to content
AnyRouter

SLM Lab

Small model on your GPU, then the cloud when it's out of its depth.

SmolLM2 135M Instructliquid/lfm-2.5-2.6b
Ask the on-device model
Nothing leaves this browser until you escalate.
0 / 400 characters

First run downloads about 114 MiB from huggingface.co and caches it here.

Answers
Local versus AnyRouter, same prompt.

Run the prompt on your GPU; the answer and its numbers appear here.

SmolLM2 135M Instruct135M params · Apache-2.0 · WebGPU114 MiB
Repo
HuggingFaceTB/SmolLM2-135M-Instruct
Quantization
q4f16
Weights
huggingface.co
Cache
Browser Cache API

How it works

A prompt card loads into a fixture. A small model cartridge runs on your GPU and a gauge reads its tokens per second and a score. When the score is low, the same prompt goes on to the cloud model.

Fig. 1on-device first
1234
  1. 1A prompt card loads into the fixture.
  2. 2The 135M cartridge runs on your GPU.
  3. 3The gauge reads tokens per second and a score.
  4. 4A low score sends the same prompt to the cloud model.
11.1 tok/s · score 0.38 · escalate