prattle_not
LatentPort's KV-cache handoff, reproduced and carried to new model pairs. Bridges, base models and the paper.
UpdatedNote The one-token bridge from SmolLM2-135M to SmolLM2-360M: 1.87x sooner to the first token, in the browser. Try it: https://agentic-underground.github.io/prattle_not/
LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay
Paper • 2609.25053 • Published • 6Note LatentPort, by Simon Villani: the method every bridge here reproduces.
HuggingFaceTB/SmolLM2-135M-Instruct
Text Generation • 0.1B • Updated • 788k • 475Note The reader: it reads the prompt, and the bridge translates its memory.
HuggingFaceTB/SmolLM2-360M-Instruct
Text Generation • 0.4B • Updated • 230k • 248Note The answerer: it answers from the translated memory, reading one token itself.
agentic-underground/qwen3-0.6b-to-1.7b-bridge
UpdatedNote The bridge from Qwen3-0.6B to Qwen3-1.7B, a second model family: held-out NCR 0.805 with the paper's correction.
Qwen/Qwen3-0.6B
Text Generation • 0.8B • Updated • 15.3M • 1.92kNote Pair B's reader: it reads the prompt, and the bridge translates its memory.
Qwen/Qwen3-1.7B
Text Generation • 2B • Updated • 2.66M • 635Note Pair B's answerer: it answers from the translated memory.
agentic-underground/smollm2-360m-to-1.7b-bridge
UpdatedNote The one-token bridge from SmolLM2-360M to SmolLM2-1.7B, a five-times size gap: held-out NCR 0.845, and 7 of 7 answered as the 1.7B.
HuggingFaceTB/SmolLM2-1.7B-Instruct
Text Generation • 2B • Updated • 240k • 773Note The larger answerer: it answers from the 360M's translated memory, reading one token itself.
agentic-underground/gemma-3-270m-to-1b-bridge
UpdatedNote The bridge from Gemma 3 270M to Gemma 3 1B, a third family: 1.91x sooner to the first token. Gated: a Gemma Model Derivative, under the Gemma Terms of Use.
google/gemma-3-270m-it
Text Generation • 0.3B • Updated • 38.3k • 722Note The Gemma pair's reader (gated, Gemma Terms of Use).
google/gemma-3-1b-it
Text Generation • 1.0B • Updated • 195k • 1.33kNote The Gemma pair's answerer (gated, Gemma Terms of Use).
agentic-underground/qwen3.5-0.8b-to-2b-bridge
UpdatedNote The bridge from Qwen3.5-0.8B to Qwen3.5-2B, a hybrid of attention and Gated DeltaNet layers: held-out NCR 0.917, and 1.36x sooner to the first token.
Qwen/Qwen3.5-0.8B
Image-Text-to-Text • 0.9B • Updated • 2.47M • 854Note The Qwen3.5 pair's reader: it reads the prompt, and its recurrent state crosses as it is.
Qwen/Qwen3.5-2B
Image-Text-to-Text • 2B • Updated • 5.38M • 495Note The Qwen3.5 pair's answerer: it answers from the translated memory, reading one token itself.
agentic-underground/llama-3.2-1b-to-3b-bridge
UpdatedNote Built with Llama. The bridge from Llama 3.2 1B to Llama 3.2 3B, across head widths 64 and 128: held-out NCR 0.965, and 1.70x sooner to the first token. Gated, under the Llama 3.2 Community License.
meta-llama/Llama-3.2-1B-Instruct
Text Generation • 1B • Updated • 1.34M • 2.01kNote The Llama pair's reader (gated, Llama 3.2 Community License).
meta-llama/Llama-3.2-3B-Instruct
Text Generation • 3B • Updated • 723k • 3.06kNote The Llama pair's answerer (gated, Llama 3.2 Community License).
agentic-underground/qwen3.5-4b-to-9b-bridge
UpdatedNote LatentPort's own pair, Qwen3.5-4B to Qwen3.5-9B, re-fitted with the paper's method: held-out NCR 0.989, and 7 of 7 answered. The speed-up is small today.
Qwen/Qwen3.5-4B-Base
Image-Text-to-Text • 5B • Updated • 497k • 133Note The paper's reader: it reads the prompt, and its recurrent state crosses as it is.
Qwen/Qwen3.5-9B-Base
Image-Text-to-Text • 10B • Updated • 267k • 163Note The paper's answerer: it answers from the translated memory, reading one token itself.
agentic-underground/lfm2-350m-to-1.2b-bridge
UpdatedNote The bridge from LFM2-350M to LFM2-1.2B, short convolutions mapped across widths: 7 of 7 answered, and 1.88x sooner. Under the LFM Open License v1.0 (USD 10M revenue threshold).
LiquidAI/LFM2-350M
Text Generation • 0.4B • Updated • 17k • 260Note The LFM2 pair's reader (LFM Open License v1.0).
LiquidAI/LFM2-1.2B
Text Generation • 1B • Updated • 17.4k • 370Note The LFM2 pair's answerer (LFM Open License v1.0).
agentic-underground/qwen3-4b-to-8b-bridge
UpdatedNote The bridge from Qwen3-4B to Qwen3-8B: held-out NCR 0.890, and 7 of 7 answered as each model alone. 1.12x sooner across two GPUs.
Qwen/Qwen3-4B
Text Generation • 4B • Updated • 3.15M • 805Note The larger Qwen3 pair's reader.
Qwen/Qwen3-8B
Text Generation • 8B • Updated • 6.19M • 2.21kNote The larger Qwen3 pair's answerer: it answers from the translated memory, reading one token itself.
agentic-underground/qwen3-0.6b-to-8b-bridge
UpdatedNote The bridge from Qwen3-0.6B to Qwen3-8B, a 13.6x gap: 7 of 7 answered as the 8B alone, held-out NCR 0.803, and 2.82x sooner to the first token on one GPU.
agentic-underground/qwen3-0.6b-to-32b-bridge
UpdatedNote The bridge from Qwen3-0.6B to a Q4K Qwen3-32B: 4.21x sooner at ~256 tokens, 7.17x at ~1K. Below our answer bar: 6 of 7 where the 32B alone gets 7.
Qwen/Qwen3-32B
Text Generation • 33B • Updated • 3.95M • • 798Note The largest answerer: it answers from a 0.6B's translated memory, quantized to Q4K.
agentic-underground/granite-4.0-h-1b-to-h-tiny-bridge
UpdatedNote The bridge from Granite 4.0 H 1B to H Tiny, a 64-expert Mamba-2 hybrid: each recurrent layer's inputs are translated and H Tiny builds its own state. 7 of 7, held-out NCR 0.924, about 2x sooner.
ibm-granite/granite-4.0-h-1b
Text Generation • 1B • Updated • 4.8k • 149Note The Granite pair's reader: its residual streams are translated into H Tiny's layer inputs.
ibm-granite/granite-4.0-h-tiny
Text Generation • 7B • Updated • 35.2k • 219Note The Granite pair's answerer, a 64-expert mixture: it builds its own Mamba-2 state from the translated inputs.
agentic-underground/granite-4.0-h-350m-to-h-small-bridge
UpdatedNote The bridge from Granite 4.0 H 350M to a Q4K H Small, about 95x larger: 7 of 7 answered as H Small alone, held-out NCR 0.767, and 1.75x sooner across two GPUs.
ibm-granite/granite-4.0-h-350m
Text Generation • 0.3B • Updated • 5.81k • 114Note Super-bridge 2's reader: its residual streams are translated into H Small's layer inputs.
ibm-granite/granite-4.0-h-small
Text Generation • 32B • Updated • 26.3k • 313Note Super-bridge 2's answerer, a 72-expert 32B quantized to Q4K: it builds its own Mamba-2 state from the translated inputs.
agentic-underground/gemma-4-e2b-to-31b-bridge
UpdatedNote The bridge from Gemma 4 E2B to a Q4K Gemma 4 31B, each layer fitted in its own shape: 7 of 7 answered as the 31B alone, held-out NCR 0.809, and 1.97x sooner.
google/gemma-4-E2B-it
Any-to-Any • 5B • Updated • 2.73M • 1.16kNote Super-bridge 3's reader, Apache-2.0: its residual streams are translated into the 31B's keys and values.
google/gemma-4-31B-it
Image-Text-to-Text • 31B • Updated • 5.97M • • 4.44kNote Super-bridge 3's answerer, quantized to Q4K: sliding and global layers, each filled in its own shape.