A fitted KV bridge: Qwen3.5-0.8B reads the prompt, and this bridge translates its internal state into the memory Qwen3.5-2B would have built. The 2B then answers from that memory, reading one token of the prompt itself.
It applies LatentPort, Simon Villani's method (arXiv:2609.25053), to a hybrid model: Qwen3.5 interleaves three linear-attention (Gated DeltaNet) layers with every full-attention layer. The reproduction it belongs to is 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.
Qwen3.5 keeps two kinds of memory, and the bridge treats them differently:
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 ridge maps alone | 0.884 | 0.058 | — |
| plus the paper's learned correction | 0.917 | 0.019 | 7, the same as each model alone |
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 2B's loss on its own native
reply, scored teacher-forced. It is a loss ratio, not a rate of matching tokens.
Both Qwen3.5 models answer all seven questions on their own, so the answer count shows the bridge does no harm but cannot show a gain. The wrong-donor column is the control: the same bridge fed another question's memory scores almost nothing. Per token, the 2B's loss on its own reply is 0.242 reading everything, 0.391 bridged with the correction, and 2.032 reading nothing.
| time to first token, bridged 2B vs the 2B reading the prompt | 1.36× (1.35–1.36) sooner |
| the ceiling in the same run (2B alone ÷ 0.8B alone) | 1.50× |
The 2B alone took 106 ms to its first token, bridged 78 ms, and the 0.8B alone 70 ms. 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 2B read no prompt history itself. The figure is the geometric mean of native ÷ bridged with its 95% interval, LatentPort's own statistic. The two models are close in size, so the ceiling is low; the bridge reaches most of it.
To repeat it from a PrattleNot checkout:
PRATTLE_DEVICES=cuda:0 just speedup models/Qwen3.5-0.8B models/Qwen3.5-2B models/bridges/qwen3.5-0.8b-to-2b.safetensors
2fc06364715b967f1860aea9cf38778875588b17 and
Qwen/Qwen3.5-2B at
15852e8c16360a2fea060d615a32b45270f8a8fc, both Apache-2.0. 24 layers each, and 2 KV heads of
width 256 in both. Only the language model is used.just bridge-fit-qwen3_5): closed-form ridge regressions (λ = 10⁻³), one per
full-attention layer for keys and one for values, read from the 0.8B's residual stream at two
taps. R² is 0.82–0.90 for keys and 0.79–0.90 for values, lowest at the last layer.just bridge-correct-qwen3_5): the paper's rank-4 residual correction per
mapped layer and role, trained through the frozen 2B against the native 2B's distribution
(451 questions, 3 epochs, loss 0.153 → 0.113 → 0.100), and folded into the file as I + BUᵀ.
The recurrent layers carry no correction.crates/core/src/bridge_corpus.rs.
The held-out questions are never fitted on.The method, and how to repeat every number, are in PrattleNot's
documents/BRIDGE.md
and documents/REPRODUCTION.md.
qwen3.5-0.8b-to-2b.safetensors, 28 MB, with no __metadata__:
| tensors | what they hold |
|---|---|
k.<l>, v.<l> (f16, 2049 × 512) |
per full-attention layer, the ridge map from the 0.8B's two residual-stream taps (plus an intercept) to the 2B's keys / values |
ck.<l>, cv.<l> (f32, 256 × 256) |
per full-attention layer, the learned correction folded to one matrix per head width |
layer_of |
which 0.8B layer feeds each 2B 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 Qwen3.5 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, the one-token handoff and the learned correction are the paper's (arXiv:2609.25053v1). Qwen3.5 is by the Qwen team, under Apache-2.0. This bridge is released under Apache-2.0.
Qwen/Qwen3.5-0.8B