PrattleNot LatentPort, in your browser
Loading the model…
LatentPort's KV-cache handoff · running locally · WebAssembly on your CPU

A small model reads your prompt. A larger one answers from its memory.

Turn the bridge on and SmolLM2-135M reads your question while SmolLM2-360M, almost three times larger, answers from the small one's translated memory, reading exactly one token of your prompt itself. Both models and the bridge run in this tab. Every claim below has a button that checks it on your machine.

1.87×
sooner to the first token than the 360M reading the prompt itself
95% CI 1.83–1.91 · CPU, WebAssembly
1 token
of your prompt is all the answering model reads; the rest crosses as memory
counted on every reply
6 / 7
held-out questions right, as many as the 360M reading everything
3 / 7 without the correction
3.39×
LatentPort's own 4B → 27B pair, reproduced on two RTX 3090s at ~16K tokens
upstream's code and statistic
Bridge off
SmolLM2-135M reads your prompt; SmolLM2-360M, almost three times larger, answers from its translated memory.
The answer appears here
Time to first token—

Who read what counted by the models, this reply

Reads
SmolLM2-135M
—
prompt tokens
Bridge
fitted maps
—
bridge tokens
Answers
SmolLM2-360M
—
prompt history read
prompt — read by the small model — read by the larger model —
Ask with the bridge on to watch the prompt cross as memory.

Your first tokens this session

Each reply's time to first token lands here.

The fair race is the 360M reading the prompt itself against the bridge:

1.87× sooner, and how it is measured

Time to first token for the bridged 360M against the 360M reading the whole prompt, on the same machine, in the same tab: 8 prompts × 3 paired repetitions, the arm order rotating, one warm-up per arm discarded.

The figure is the geometric mean of native ÷ bridged over prompts with its 95% interval, LatentPort's own statistic. The bridged time includes the 135M reading the prompt.

One bridge token

LatentPort's handoff: the larger model reads exactly one token of the prompt itself, and everything before it arrives as translated key/value memory.

The page reads the counts from the models themselves on every reply: prompt tokens, tokens the 135M read, prompt history the 360M read, and bridge tokens.

6 of 7 held-out questions

Seven questions never used in fitting, each with one checkable answer. Bridged, the 360M gets 6 right, as many as the 360M reading everything. With the ridge maps alone, before the paper's learned correction, it kept only the shape of an answer: 3 of 7.

Fidelity on held-out replies: NCR 0.930. The paper's corrected bridge, on its own pair: 0.918.

The model card →

LatentPort's own pair, reproduced

Qwen3.5-4B → 27B, timed with upstream's own runtime, timing code and statistic, on two RTX 3090s: 1.36× sooner at 128 prompt tokens, 2.94× at ~2K and 3.39× at ~16K.

The full tables, every run, and a guide to repeat them are in the repository.

The result →

Time to first token

From pressing Ask to the first token. With the bridge off, the 135M answers alone; with it on, the 360M answers from the 135M's memory, so these two times compare different models. The fair comparison is on the Measure tab.

The counters

Reported by the models after each reply, not computed by the page: how many prompt tokens the reading model read, how much prompt history the answering model read, and how many bridge tokens it read itself.

Bridged, the 135M reads all but the last token, the 360M reads none of the prompt, and one bridge token crosses.

Why these controls

An answer that looks right could come from the one token the 360M reads, not the memory. So the same question runs with a different question's memory: if the answer changes, it was reading the memory.

Restoring the 135M's own memory must reproduce its reply exactly, which shows the memory is moved without loss. Each run is capped at 32 tokens.

NCR, the native-closure ratio

(ℓ_empty − ℓ_bridged) / (ℓ_empty − ℓ_native), the paper's Eq. 8: each ℓ is the larger model's loss on its own native reply. 1 means the bridged model is as sure of that reply as when it reads everything.

Read it beside each card's wrong-memory control: where an empty model is very far off, any memory scores high.