File size: 14,735 Bytes
b503124
 
f0c89ea
 
 
b503124
 
 
 
9e8a37f
f0c89ea
9e8a37f
f0c89ea
9e8a37f
 
 
64e4100
 
9e8a37f
a61cf69
9e8a37f
 
 
 
 
e2edb82
 
 
 
64e4100
9e8a37f
64e4100
9e8a37f
 
 
 
 
 
 
64e4100
9e8a37f
 
 
 
 
f0c89ea
 
9e8a37f
 
 
 
 
f0c89ea
64e4100
9e8a37f
 
 
 
 
 
 
 
64e4100
9e8a37f
71bf51a
 
 
 
 
 
 
 
 
 
 
 
 
ae12dba
 
 
 
 
 
 
 
 
 
 
dedade8
 
3fb94b5
dedade8
 
 
 
 
 
 
 
a93374f
cb0f440
 
 
 
 
a93374f
cb0f440
 
 
 
9204c81
 
 
 
 
 
 
 
 
 
d34f468
9204c81
8880537
 
 
 
 
 
cb0f440
8880537
 
 
 
 
 
34abae0
 
 
 
 
 
 
71bf51a
34abae0
 
 
 
 
 
 
0f159dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3fb94b5
 
 
 
 
 
 
 
 
 
 
 
 
0fcf800
 
 
 
 
 
 
 
 
 
 
 
14037f6
 
 
 
 
 
 
 
 
 
 
01220d9
14037f6
2248f68
 
 
 
 
 
 
 
 
 
 
d34f468
2248f68
d2f0ab5
9e8a37f
d2f0ab5
 
 
 
 
 
 
 
d34f468
d2f0ab5
 
 
9e8a37f
 
f0c89ea
 
 
c7886eb
 
9e8a37f
 
 
f0c89ea
64e4100
9e8a37f
8880537
34abae0
0f159dc
 
9e8a37f
 
 
 
 
 
f0c89ea
9e8a37f
64e4100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
---
title: README
emoji: πŸ›Έ
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false
---

<div align="center">

# πŸ›Έ agentic-underground

**Open tools and research for building software with AI agents** β€”
disciplined workflows, safety, token efficiency, and reproducible model research.

<a href="https://huggingface.co/collections/agentic-underground/prattle-not-6af7ac5e95e0fc80de563d61"><img alt="Collection: prattle_not" src="https://img.shields.io/badge/collection-prattle__not-6366f1"></a>
<a href="https://agentic-underground.github.io/prattle_not/"><img alt="Live demo in your browser" src="https://img.shields.io/badge/demo-runs%20in%20your%20browser-a855f7"></a>
<a href="https://github.com/agentic-underground"><img alt="Source on GitHub" src="https://img.shields.io/badge/source-GitHub-24292f?logo=github"></a>
<a href="https://huggingface.co/papers/2609.25053"><img alt="Builds on LatentPort by Simon Villani (arXiv 2609.25053)" src="https://img.shields.io/badge/builds%20on-LatentPort%20by%20Simon%20Villani-b31b1b"></a>

</div>

---

> **⚠️ 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 β€” **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 β†’](https://huggingface.co/collections/agentic-underground/prattle-not-6af7ac5e95e0fc80de563d61)**

### ✨ Featured β€” [`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.

| | |
|:--|:--|
| ⚑ **Time to first token** | **1.87Γ—** sooner (95% CI 1.83–1.91), in WebAssembly, on a CPU |
| πŸͺ™ **Bridge tokens** | **1** β€” the 360M reads one token of the prompt itself |
| 🎯 **Held-out questions** | **6 of 7** correct β€” as many as the 360M reading everything |
| πŸ” **Fidelity (NCR)** | **0.930** β€” the paper's corrected bridge on its own pair: 0.918 |

▢️ **[Try it in your browser](https://agentic-underground.github.io/prattle_not/)** β€” both models and the
bridge download once and run locally. Turn the bridge on, ask anything, and watch the counters show
who read what.

### πŸ“ The reproduction behind it

Before carrying the method to new models, we reproduced LatentPort's own **4B β†’ 27B** pair on two
RTX 3090s with upstream's timing code and statistic: **1.36Γ—** sooner at 128 prompt tokens, rising to
**2.94Γ—** at ~2K and **3.39Γ—** 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).

### πŸ“„ The paper's own pair β€” [`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.

| | |
|:--|:--|
| πŸ” **Fidelity (NCR)** | **0.989** with the paper's learned correction (0.974 from the maps alone), our highest yet |
| πŸ§ͺ **Control (wrong donor)** | 0.308, high because the recurrent state carries general context whichever prompt made it |
| 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
| ⚑ **Time to first token** | **1.16Γ—** sooner (95% CI 1.16–1.16) on two RTX 3090s at ~256 tokens; ceiling 1.38Γ—. Small: the two models share a recurrent core that costs the same in both |

### 🧬 Qwen3 at a larger size β€” [`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.

| | |
|:--|:--|
| πŸ” **Fidelity (NCR)** | **0.890** with the paper's learned correction |
| πŸ§ͺ **Control (wrong donor)** | βˆ’0.026 |
| 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
| ⚑ **Time to first token** | **1.12Γ—** sooner (95% CI 1.12–1.12) on two RTX 3090s at ~256 tokens; ceiling 1.61Γ— |

### πŸš€ Fifty-four times larger β€” [`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.

| | |
|:--|:--|
| ⚑ **Time to first token** | **4.21Γ—** sooner (95% CI 4.19–4.23) at ~256 tokens, **7.17Γ—** at ~1K, on two RTX 3090s; ceilings 11.60Γ— and 14.33Γ— |
| 🎯 **Held-out questions** | **6 of 7** β€” the 32B alone gets 7, so it is **below our answer bar** |
| πŸ” **Fidelity (NCR)** | **0.866** with the paper's learned correction (0.794 from the maps alone) |
| πŸ§ͺ **Control (wrong donor)** | 0.197 |

### πŸͺœ A 13.6Γ— gap, at the bar β€” [`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.

| | |
|:--|:--|
| ⚑ **Time to first token** | **2.82Γ—** sooner (95% CI 2.82–2.83) on one RTX 3090 at ~256 tokens; ceiling 6.03Γ— |
| 🎯 **Held-out questions** | **7 of 7**, the same as the 8B alone |
| πŸ” **Fidelity (NCR)** | **0.803** with the paper's learned correction |
| πŸ§ͺ **Control (wrong donor)** | 0.092 |

### πŸ“ A larger gap β€” [`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.

| | |
|:--|:--|
| 🎯 **Held-out questions** | **7 of 7**, as the 1.7B reading everything (the 360M alone: 6) |
| πŸ” **Fidelity (NCR)** | **0.845** with the paper's learned correction (0.611 from the maps alone) |
| πŸ§ͺ **Control (wrong donor)** | βˆ’0.266 |
| ⚑ **Time to first token** | **1.67Γ—** sooner (95% CI 1.66–1.69) on one RTX 3090 at ~256 tokens; ceiling 2.59Γ— |

### πŸ’Ž A third family β€” [`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.

| | |
|:--|:--|
| ⚑ **Time to first token** | **1.91Γ—** sooner (95% CI 1.86–1.97) on one RTX 3090 at ~256 tokens; ceiling 2.79Γ— |
| πŸ” **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 |
| 🎯 **Held-out questions** | **6 of 7**, the same as the 270M alone (the 1B: 7) β€” no answer gain yet |

βš–οΈ **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).

### πŸ¦™ A fourth family β€” [`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).

| | |
|:--|:--|
| πŸ” **Fidelity (NCR)** | **0.965** with the paper's learned correction (0.818 from the maps alone) |
| πŸ§ͺ **Control (wrong donor)** | βˆ’0.040 |
| 🎯 **Held-out questions** | **6 of 7**, the same as the 3B alone (the 1B: 7). No answer gain yet |
| ⚑ **Time to first token** | **1.70Γ—** sooner (95% CI 1.66–1.74) on one RTX 3090 at ~256 tokens; ceiling 2.54Γ— |

βš–οΈ **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).

### 🌊 Short convolutions β€” [`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.

| | |
|:--|:--|
| ⚑ **Time to first token** | **1.88Γ—** sooner (95% CI 1.86–1.91) on one RTX 3090 at ~256 tokens; ceiling 2.50Γ— |
| 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
| πŸ” **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 |

βš–οΈ **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.

### 🐘 Ninety-five times larger, through Mamba-2 β€” [`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.

| | |
|:--|:--|
| 🎯 **Held-out questions** | **7 of 7**, the same as H Small alone (the 350M alone gets 5) |
| πŸ” **Fidelity (NCR)** | **0.767** from the maps alone |
| πŸ§ͺ **Control (wrong donor)** | 0.161 |
| ⚑ **Time to first token** | **1.75Γ—** sooner (95% CI 1.74–1.75) on two RTX 3090s at ~256 tokens, **1.62Γ—** at ~1K; ceilings 7.30Γ— and 7.58Γ—. H Small rebuilds 36 Mamba-2 states at the handoff |

### πŸ’Ž Gemma 4, two layer shapes β€” [`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 Γ— 256) and 10
global layers (4 Γ— 512) are each fitted in their own shape. Apache-2.0, not gated.

| | |
|:--|:--|
| 🎯 **Held-out questions** | **7 of 7**, the same as the 31B alone |
| πŸ” **Fidelity (NCR)** | **0.809** from the maps alone |
| πŸ§ͺ **Control (wrong donor)** | 0.448 |
| ⚑ **Time to first token** | **1.97Γ—** sooner (95% CI 1.94–2.00) on two RTX 3090s at ~256 tokens; ceiling 5.10Γ—. Longer prompts await an attention path that fits a 24 GB card |

### 🐍 Mamba-2, translated β€” [`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.

| | |
|:--|:--|
| 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
| πŸ” **Fidelity (NCR)** | **0.924** with the paper's learned correction (0.916 from the maps alone) |
| πŸ§ͺ **Control (wrong donor)** | βˆ’0.175 |
| ⚑ **Time to first token** | **2.06Γ—** sooner (95% CI 1.95–2.18) on one RTX 3090 at ~256 tokens; ceiling 3.95Γ—. Its experts route on the host CPU, so a busy CPU lowers it |

### πŸ”€ A hybrid model β€” [`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.

| | |
|:--|:--|
| πŸ” **Fidelity (NCR)** | **0.917** with the paper's learned correction (0.884 from the maps alone) |
| πŸ§ͺ **Control (wrong donor)** | 0.019 |
| 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
| ⚑ **Time to first token** | **1.36Γ—** sooner (95% CI 1.35–1.36) on one RTX 3090 at ~256 tokens; ceiling 1.50Γ— |

### 🧬 A second family β€” [`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.

| | |
|:--|:--|
| πŸ” **Fidelity (NCR)** | **0.805** with the paper's learned correction (0.645 from the maps alone) |
| 🎯 **Held-out questions** | **7 of 7** correct, as many as the 1.7B reading everything |
| πŸ§ͺ **Control (wrong donor)** | βˆ’0.127: the bridge fed another question's memory scores far lower |
| ⚑ **Time to first token** | **1.47Γ—** sooner (95% CI 1.46–1.48) on one RTX 3090 at ~256 tokens; ceiling 2.03Γ— |

Each new pair gets its own repository here and joins the collection once its weights are fitted
and measured.

---

## How we publish

- πŸ”¬ **Measured** β€” 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.
- πŸ“Œ **Pinned** β€” our pages fetch weights by commit, never by `main`, so a later upload cannot change
  what a published result measured.
- 🧩 **Our own work only** β€” we host what we fitted. Base models stay with their publishers and are
  linked, never mirrored.
- πŸ™ **Attributed** β€” 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 β†’ 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).