Nanbeige 4.2–3B local, private, yours

Hey. What’s on your mind?

A little model, right on your device.
Your conversation stays in this tab.

Message NanbeigeENTER TO SEND · SHIFT + ENTER FOR A NEW LINE

Answer limit

Ready when you are

Nanbeige4.2–3B2.35 GB default download · saved in your browser when possibleModel settings +

THE MODEL

Nanbeige4.2–3B

ON DEMAND

Model profile

EngineContext

Compatibility modeNo subgroups · compatibility profile

First download: about 2.35 GB. Cached when browser storage allows.

THE SPEEDOMETER LOCAL

How fast is yours?

Every answer is a benchmark.
These numbers come from your device.

—

tokens / secondGeneration speed

—Time to first token

—Prompt tokens

—Output tokens

—Total run time

Run your first prompt to get a reading. Loading time is measured separately.

Session history

No runs yet. Your GPU is waiting.

JSON includes prompts, output, per-token timings, and device details. Saved only when you export.

Take a closer lookRepeatable benchmark & validation suite

One stack. Two passes.

22 shared transformer layers run twice, with separate attention history for each pass. The trained two-pass configuration stays intact.

22 LAYERS↻2 PASSES

Q4 weights, FP32 activations, and 44 separate KV-cache slots in the custom engine.

WebGPU is required. A user-reported iPad Pro M2 run reached 18.0 tok/s. Phones still need physical-device testing; memory limits vary. Start with 512 context; larger contexts use more GPU memory. Results vary with browser, temperature, and other running apps.