--- title: README emoji: ๐ธ colorFrom: indigo colorTo: purple sdk: static pinned: false ---
--- > **โ ๏ธ Security note for every bridge here.** A bridged model does not read the prompt itself, the system > prompt included, and rules set in a system prompt are broken more often bridged (measured on six pairs). > Do not rely on system-prompt instructions as guardrails for a bridged model; enforce policy outside it. ## On the Hub: the prattle_not family **prattle_not** reproduces [**LatentPort**](https://huggingface.co/papers/2609.25053), Simon Villani's method for handing a conversation from one language model to another *through its memory*. A small model reads the prompt; a fitted **bridge** translates its key/value cache into the memory a larger model would have built; the larger model answers from it โ **without reading the prompt itself**, so its first token arrives sooner. What we publish here are the **bridges**: the fitted weights that make that handoff work between a specific pair of models. Everything is gathered in the **[prattle_not collection โ](https://huggingface.co/collections/agentic-underground/prattle-not-6af7ac5e95e0fc80de563d61)** ### โจ Featured โ [`smollm2-135m-to-360m-bridge`](https://huggingface.co/agentic-underground/smollm2-135m-to-360m-bridge) SmolLM2-135M reads; SmolLM2-360M, almost three times larger, answers. Two models with different shapes, bridged with the paper's **one-token handoff** and its learned correction. | | | |:--|:--| | โก **Time to first token** | **1.87ร** sooner (95% CI 1.83โ1.91), in WebAssembly, on a CPU | | ๐ช **Bridge tokens** | **1** โ the 360M reads one token of the prompt itself | | ๐ฏ **Held-out questions** | **6 of 7** correct โ as many as the 360M reading everything | | ๐ **Fidelity (NCR)** | **0.930** โ the paper's corrected bridge on its own pair: 0.918 | โถ๏ธ **[Try it in your browser](https://agentic-underground.github.io/prattle_not/)** โ both models and the bridge download once and run locally. Turn the bridge on, ask anything, and watch the counters show who read what. ### ๐ The reproduction behind it Before carrying the method to new models, we reproduced LatentPort's own **4B โ 27B** pair on two RTX 3090s with upstream's timing code and statistic: **1.36ร** sooner at 128 prompt tokens, rising to **2.94ร** at ~2K and **3.39ร** at ~16K. The full tables, every run, and a step-by-step guide to repeat them are in the [prattle_not repository](https://github.com/agentic-underground/prattle_not). ### ๐ The paper's own pair โ [`qwen3.5-4b-to-9b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-4b-to-9b-bridge) Qwen3.5-4B reads; Qwen3.5-9B answers. This is the pair Simon Villani's paper fits, re-fitted independently with its method. The attention layers' keys and values are mapped, and the 24 recurrent layers cross as the 4B's own state. | | | |:--|:--| | ๐ **Fidelity (NCR)** | **0.989** with the paper's learned correction (0.974 from the maps alone), our highest yet | | ๐งช **Control (wrong donor)** | 0.308, high because the recurrent state carries general context whichever prompt made it | | ๐ฏ **Held-out questions** | **7 of 7**, the same as each model alone | | โก **Time to first token** | **1.16ร** sooner (95% CI 1.16โ1.16) on two RTX 3090s at ~256 tokens; ceiling 1.38ร. Small: the two models share a recurrent core that costs the same in both | ### ๐งฌ Qwen3 at a larger size โ [`qwen3-4b-to-8b-bridge`](https://huggingface.co/agentic-underground/qwen3-4b-to-8b-bridge) Qwen3-4B reads; Qwen3-8B answers, across all 36 layers. | | | |:--|:--| | ๐ **Fidelity (NCR)** | **0.890** with the paper's learned correction | | ๐งช **Control (wrong donor)** | โ0.026 | | ๐ฏ **Held-out questions** | **7 of 7**, the same as each model alone | | โก **Time to first token** | **1.12ร** sooner (95% CI 1.12โ1.12) on two RTX 3090s at ~256 tokens; ceiling 1.61ร | ### ๐ Fifty-four times larger โ [`qwen3-0.6b-to-32b-bridge`](https://huggingface.co/agentic-underground/qwen3-0.6b-to-32b-bridge) Qwen3-0.6B reads; Qwen3-32B, quantized to Q4K, answers. Published with its miss stated. | | | |:--|:--| | โก **Time to first token** | **4.21ร** sooner (95% CI 4.19โ4.23) at ~256 tokens, **7.17ร** at ~1K, on two RTX 3090s; ceilings 11.60ร and 14.33ร | | ๐ฏ **Held-out questions** | **6 of 7** โ the 32B alone gets 7, so it is **below our answer bar** | | ๐ **Fidelity (NCR)** | **0.866** with the paper's learned correction (0.794 from the maps alone) | | ๐งช **Control (wrong donor)** | 0.197 | ### ๐ช A 13.6ร gap, at the bar โ [`qwen3-0.6b-to-8b-bridge`](https://huggingface.co/agentic-underground/qwen3-0.6b-to-8b-bridge) Qwen3-0.6B reads; Qwen3-8B, 13.6 times larger, answers. | | | |:--|:--| | โก **Time to first token** | **2.82ร** sooner (95% CI 2.82โ2.83) on one RTX 3090 at ~256 tokens; ceiling 6.03ร | | ๐ฏ **Held-out questions** | **7 of 7**, the same as the 8B alone | | ๐ **Fidelity (NCR)** | **0.803** with the paper's learned correction | | ๐งช **Control (wrong donor)** | 0.092 | ### ๐ A larger gap โ [`smollm2-360m-to-1.7b-bridge`](https://huggingface.co/agentic-underground/smollm2-360m-to-1.7b-bridge) SmolLM2-360M reads; SmolLM2-1.7B, about five times larger and shaped differently, answers from **one bridge token**. Here the two models differ on the answer test, so it can show a gain. | | | |:--|:--| | ๐ฏ **Held-out questions** | **7 of 7**, as the 1.7B reading everything (the 360M alone: 6) | | ๐ **Fidelity (NCR)** | **0.845** with the paper's learned correction (0.611 from the maps alone) | | ๐งช **Control (wrong donor)** | โ0.266 | | โก **Time to first token** | **1.67ร** sooner (95% CI 1.66โ1.69) on one RTX 3090 at ~256 tokens; ceiling 2.59ร | ### ๐ A third family โ [`gemma-3-270m-to-1b-bridge`](https://huggingface.co/agentic-underground/gemma-3-270m-to-1b-bridge) Gemma 3 270M reads; Gemma 3 1B answers, across sliding-window and global attention layers. | | | |:--|:--| | โก **Time to first token** | **1.91ร** sooner (95% CI 1.86โ1.97) on one RTX 3090 at ~256 tokens; ceiling 2.79ร | | ๐ **Fidelity (NCR)** | **0.870**, but a wrong prompt's memory reaches 0.600 here: an empty Gemma 3 1B is far off, so read the two together | | ๐ฏ **Held-out questions** | **6 of 7**, the same as the 270M alone (the 1B: 7) โ no answer gain yet | โ๏ธ **Gated, and not Apache-2.0:** it is a Gemma Model Derivative, distributed under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). ### ๐ฆ A fourth family โ [`llama-3.2-1b-to-3b-bridge`](https://huggingface.co/agentic-underground/llama-3.2-1b-to-3b-bridge) **Built with Llama.** Llama 3.2 1B reads; Llama 3.2 3B answers, across different head widths (64 and 128). | | | |:--|:--| | ๐ **Fidelity (NCR)** | **0.965** with the paper's learned correction (0.818 from the maps alone) | | ๐งช **Control (wrong donor)** | โ0.040 | | ๐ฏ **Held-out questions** | **6 of 7**, the same as the 3B alone (the 1B: 7). No answer gain yet | | โก **Time to first token** | **1.70ร** sooner (95% CI 1.66โ1.74) on one RTX 3090 at ~256 tokens; ceiling 2.54ร | โ๏ธ **Gated, and not Apache-2.0:** it is a derivative work of Llama 3.2, distributed under the [Llama 3.2 Community License](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/blob/main/LICENSE.txt). ### ๐ Short convolutions โ [`lfm2-350m-to-1.2b-bridge`](https://huggingface.co/agentic-underground/lfm2-350m-to-1.2b-bridge) LFM2-350M reads; LFM2-1.2B answers. Liquid AI's LFM2 mixes short convolutions with attention, and the two sizes differ in width, so the convolution state cannot be copied: the bridge maps it too. | | | |:--|:--| | โก **Time to first token** | **1.88ร** sooner (95% CI 1.86โ1.91) on one RTX 3090 at ~256 tokens; ceiling 2.50ร | | ๐ฏ **Held-out questions** | **7 of 7**, the same as each model alone | | ๐ **Fidelity (NCR)** | **0.978** (0.982 from the maps alone), but a wrong prompt's memory reaches 0.730: an empty LFM2 is far off, so read the two together | โ๏ธ **Gated, and not Apache-2.0:** it is distributed under the base models' [LFM Open License v1.0](https://huggingface.co/agentic-underground/lfm2-350m-to-1.2b-bridge/blob/main/LICENSE), which does not license Commercial Use by an entity with USD 10M or more in annual revenue. ### ๐ Ninety-five times larger, through Mamba-2 โ [`granite-4.0-h-350m-to-h-small-bridge`](https://huggingface.co/agentic-underground/granite-4.0-h-350m-to-h-small-bridge) Granite 4.0 H 350M reads; H Small, a 72-expert 32B quantized to Q4K, answers from translated Mamba-2 inputs. The paper's correction scores higher but answers 6 of 7, so the uncorrected maps are published. | | | |:--|:--| | ๐ฏ **Held-out questions** | **7 of 7**, the same as H Small alone (the 350M alone gets 5) | | ๐ **Fidelity (NCR)** | **0.767** from the maps alone | | ๐งช **Control (wrong donor)** | 0.161 | | โก **Time to first token** | **1.75ร** sooner (95% CI 1.74โ1.75) on two RTX 3090s at ~256 tokens, **1.62ร** at ~1K; ceilings 7.30ร and 7.58ร. H Small rebuilds 36 Mamba-2 states at the handoff | ### ๐ Gemma 4, two layer shapes โ [`gemma-4-e2b-to-31b-bridge`](https://huggingface.co/agentic-underground/gemma-4-e2b-to-31b-bridge) Gemma 4 E2B reads; Gemma 4 31B, quantized to Q4K, answers. Its 50 sliding layers (16 ร 256) and 10 global layers (4 ร 512) are each fitted in their own shape. Apache-2.0, not gated. | | | |:--|:--| | ๐ฏ **Held-out questions** | **7 of 7**, the same as the 31B alone | | ๐ **Fidelity (NCR)** | **0.809** from the maps alone | | ๐งช **Control (wrong donor)** | 0.448 | | โก **Time to first token** | **1.97ร** sooner (95% CI 1.94โ2.00) on two RTX 3090s at ~256 tokens; ceiling 5.10ร. Longer prompts await an attention path that fits a 24 GB card | ### ๐ Mamba-2, translated โ [`granite-4.0-h-1b-to-h-tiny-bridge`](https://huggingface.co/agentic-underground/granite-4.0-h-1b-to-h-tiny-bridge) Granite 4.0 H 1B reads; H Tiny, a 64-expert mixture, answers. Mamba-2 state cannot be copied between separately trained sizes, so the bridge translates each recurrent layer's **inputs**, and H Tiny builds its own state from them. | | | |:--|:--| | ๐ฏ **Held-out questions** | **7 of 7**, the same as each model alone | | ๐ **Fidelity (NCR)** | **0.924** with the paper's learned correction (0.916 from the maps alone) | | ๐งช **Control (wrong donor)** | โ0.175 | | โก **Time to first token** | **2.06ร** sooner (95% CI 1.95โ2.18) on one RTX 3090 at ~256 tokens; ceiling 3.95ร. Its experts route on the host CPU, so a busy CPU lowers it | ### ๐ A hybrid model โ [`qwen3.5-0.8b-to-2b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-0.8b-to-2b-bridge) Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet layers: the bridge maps the attention layers' keys and values, and the recurrent state crosses as the 0.8B's own. | | | |:--|:--| | ๐ **Fidelity (NCR)** | **0.917** with the paper's learned correction (0.884 from the maps alone) | | ๐งช **Control (wrong donor)** | 0.019 | | ๐ฏ **Held-out questions** | **7 of 7**, the same as each model alone | | โก **Time to first token** | **1.36ร** sooner (95% CI 1.35โ1.36) on one RTX 3090 at ~256 tokens; ceiling 1.50ร | ### ๐งฌ A second family โ [`qwen3-0.6b-to-1.7b-bridge`](https://huggingface.co/agentic-underground/qwen3-0.6b-to-1.7b-bridge) Qwen3-0.6B reads; Qwen3-1.7B answers. This is the real test of whether Simon Villani's method carries beyond one family, fitted with the same tools and measured on held-out questions. | | | |:--|:--| | ๐ **Fidelity (NCR)** | **0.805** with the paper's learned correction (0.645 from the maps alone) | | ๐ฏ **Held-out questions** | **7 of 7** correct, as many as the 1.7B reading everything | | ๐งช **Control (wrong donor)** | โ0.127: the bridge fed another question's memory scores far lower | | โก **Time to first token** | **1.47ร** sooner (95% CI 1.46โ1.48) on one RTX 3090 at ~256 tokens; ceiling 2.03ร | Each new pair gets its own repository here and joins the collection once its weights are fitted and measured. --- ## How we publish - ๐ฌ **Measured** โ every number on a card says how it was measured, and the speed-ups can be measured again on your own hardware with the published runtime. - ๐ **Pinned** โ our pages fetch weights by commit, never by `main`, so a later upload cannot change what a published result measured. - ๐งฉ **Our own work only** โ we host what we fitted. Base models stay with their publishers and are linked, never mirrored. - ๐ **Attributed** โ prattle_not is a derivative work of [LatentPort](https://latentport.com/) by Simon Villani ([source](https://github.com/LatentPort/latentport), Apache-2.0). Our bridges are Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms), and those fitted from Llama, which are Built with Llama, gated and distributed under the Llama 3.2 Community License, and those fitted from LFM2, gated and distributed under the LFM Open License v1.0. ## Elsewhere The **source** for everything here, and our other tools for building software with agents (Claude Code plugins, a static auditor for agent skills, a token-aware scheduler), live on **[GitHub โ agentic-underground](https://github.com/agentic-underground)**. Questions and corrections are welcome: open a discussion on the model concerned, or an [issue on prattle_not](https://github.com/agentic-underground/prattle_not/issues).