A fitted KV bridge: Qwen3-0.6B reads the prompt, and this bridge translates its internal state into the key/value memory Qwen3-1.7B would have built. The 1.7B then answers from that memory instead of reading the prompt itself.
It applies LatentPort, Simon Villani's method (arXiv:2609.25053), to a second model family. The reproduction it belongs to is PrattleNot. Its first bridge, SmolLM2-135M → 360M, runs in the browser. This one runs on the server for now.
⚠️ 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.
Held-out questions, never fitted on:
| share of the read-nothing → read-everything gap the bridge closes (NCR) | the same, from a wrong donor | correct of 7 | |
|---|---|---|---|
| the maps alone | 0.645 | −0.374 | 7 |
| plus the paper's learned correction | 0.805 | −0.127 | 7, the same as the 1.7B reading everything |
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 1.7B's loss on its own native
reply, scored teacher-forced. It is the share of the gap between the 1.7B reading nothing (0) and
reading the whole prompt (1) that the bridge closes. It is a loss ratio, not a rate of matching tokens.
Both Qwen3 models answer all seven questions on their own. So the answer count shows the bridge does no harm, but it can't show a gain, and NCR is the measure that tells the bridges apart. The correction closes 45% of what the maps leave. The wrong-donor column is the control: the same bridge fed another question's memory, which should score far lower.
| time to first token, bridged 1.7B vs the 1.7B reading the prompt | 1.47× (1.46–1.48) sooner |
| the ceiling in the same run (1.7B alone ÷ 0.6B alone) | 2.03× |
Measured on one RTX 3090, with both models on the same card: synthetic prompts of about 256 tokens, 8 prompts × 3 repetitions in rotated order, one bridge token, and every run's counters checked, so the bridged 1.7B read no prompt history itself. The figure is the geometric mean of native ÷ bridged with its 95% interval, LatentPort's own statistic. The speed-up grows with prompt length: on LatentPort's own 4B → 27B pair, reproduced with upstream's harness, it is 1.36× at 128 tokens and 3.39× at about 16K. So these short prompts are the hard case.
To repeat it from a PrattleNot checkout:
PRATTLE_DEVICES=cuda:0 just speedup models/Qwen3-0.6B models/Qwen3-1.7B models/bridges/qwen3-0.6b-to-1.7b.safetensors
The method, and how to repeat every number on your own machine, are in PrattleNot's
documents/BRIDGE.md
and documents/REPRODUCTION.md §5.1.
c1899de289a04d12100db370d81485cdf75e47ca and Qwen/Qwen3-1.7B
at 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e. Both have 28 layers and 8 KV heads × 128, the
matched geometry the paper assumes. Prompts use Qwen3's chat template with thinking off.qwen3-0.6b-to-1.7b.safetensors, 239 MB, with no __metadata__:
| tensors | what they hold |
|---|---|
k.<l>, v.<l> (f16, 2049 × 1024) |
per target layer, the map from the 0.6B's residual stream to the 1.7B's keys / values |
ck.<l>, cv.<l> (f32, 128 × 128) |
per target layer, the learned correction folded to one matrix per head width |
layer_of |
which 0.6B layer feeds each 1.7B layer |
meta |
the shape constants the reader needs |
It is read by prattle_core::bridge in PrattleNot, in Rust. It is not a transformers model, and it
contains no weights of either Qwen3 model. Fetch those from Qwen.
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 bridge design and its learned correction are the paper's (arXiv:2609.25053v1). Qwen3 is by the Qwen team, under Apache-2.0. This bridge is released under Apache-2.0.
Qwen/Qwen3-0.6B