A fitted KV bridge: SmolLM2-135M-Instruct reads the prompt, and this bridge translates its internal state into the key/value memory SmolLM2-360M-Instruct would have built. The 360M then answers from that memory without reading the prompt itself, so its first token arrives sooner.
It is the browser proof of PrattleNot, a reproduction of LatentPort (arXiv:2609.25053v1) carried to a new pair of models. Try it, in your browser, with nothing to install: https://agentic-underground.github.io/PrattleNot/
⚠️ Security note. When bridged, the answering model does not read the prompt itself, the system prompt included, and the bridge does not reliably carry its instructions. Measured on six of our pairs, a rule set in a system prompt is broken more often bridged than when the large model reads the prompt. Do not rely on system-prompt instructions as guardrails for a bridged model: have the large model read the system prompt itself, or enforce policy outside the model.
| time to first token, bridged 360M vs the 360M reading the prompt | 1.87× (1.83–1.91) sooner, on a 37-token prompt, in WebAssembly on the CPU |
| the ceiling (360M alone ÷ 135M alone) | 2.71× |
| bridge tokens the 360M reads | 1 — the paper's one-token handoff |
| held-out questions answered correctly | 6 of 7, the same as the 360M reading everything |
| share of the read-nothing → read-everything gap the bridge closes (NCR) | 0.930 (uncorrected: 0.802; the paper's corrected bridge on its own pair: 0.918) |
NCR is the paper's native-closure ratio (arXiv:2609.25053, Eq. 8):
(ℓ_empty − ℓ_bridged) / (ℓ_empty − ℓ_native), where each ℓ is the 360M's loss on its own native
reply, scored teacher-forced. It is the share of the gap between the 360M reading nothing (0) and
reading the whole prompt (1) that the bridge closes. It is a loss ratio, not a rate of matching tokens.
The speed-up is native ÷ bridged, the geometric mean over prompts with its 95% interval. The method,
and how to repeat every number on your own machine, is in PrattleNot's
documents/BRIDGE.md
and documents/REPRODUCTION.md.
just bridge-fit): closed-form ridge regressions, one per target layer for keys
(de-rotated) and one for values, from the residual streams of two consecutive 135M layers at the
proportionally placed depth. The 135M and the 360M have different head counts and widths, so the
maps read the residual stream rather than the 135M's own keys and values.just bridge-correct): the paper's rank-4 residual correction,
X′ = X + (XB)Uᵀ per target layer and role (16,384 parameters), trained by AdamW through the
frozen 360M against the native 360M's distribution, and folded into the file as I + BUᵀ.crates/core/src/bridge_corpus.rs.
The held-out questions in the same file are never fitted on, and a unit test keeps the two sets
disjoint.smollm2-135m-to-360m.safetensors, 48 MB, with no __metadata__:
| tensors | what they hold |
|---|---|
k.<l>, v.<l> (f16) |
per target layer, the ridge map from the 135M's residual streams to the 360M's keys / values |
ck.<l>, cv.<l> (f32) |
per target layer, the learned correction folded to a 64 × 64 matrix |
layer_of |
which 135M layer feeds each 360M layer |
meta |
the shape constants the reader needs |
It is read by prattle_core::bridge in PrattleNot, in Rust, natively or in WebAssembly. It is not a
transformers model, and it contains no weights of either SmolLM2 model — fetch those from
HuggingFaceTB.
PrattleNot is a derivative work of LatentPort by Simon Villani — https://latentport.com/, https://github.com/LatentPort/latentport, licensed under the Apache License, Version 2.0. The one-token handoff and its learned correction are the paper's (arXiv:2609.25053v1, §1.5). SmolLM2 is by Hugging Face, under Apache-2.0. This bridge is released under Apache-2.0.
HuggingFaceTB/SmolLM2-135M-Instruct