🍷 Winery LFM2.5 1.2B · Liquid Grand Cru
A Breadcrumbs merge of five LFM2.5-1.2B fine-tunes on top of LiquidAI's Instruct, blended with the Winery fusion compiler. It's 1.17B params and 1.25 GB at Q8_0. LFM2.5's hybrid conv/attention design makes it one of the fastest models in its class on phone CPUs.
Recipe
Breadcrumbs (density 0.85) over LiquidAI/LFM2.5-1.2B-Instruct, norms averaged linearly:
| Donor | Weight |
|---|---|
| huihui-ai/Huihui-LFM2.5-1.2B-Instruct-abliterated | 1.00 |
| LiquidAI/LFM2.5-1.2B-Base | 0.80 |
| yasserrmd/GLM4.7-Distill-LFM2.5-1.2B | 0.55 |
| DavidAU/LFM2.5-1.2B-Instruct-Thinking-Claude-High-Reasoning | 0.50 |
| DavidAU/LFM2.5-1.2B-MEGABRAIN-Thinking-Claude-Polaris-Deepseek-GLM | 0.45 |
name Liquid Grand Cru
out q8_0
default breadcrumbs 0.85 1.0,0.8,0.55,0.5,0.45
tensor *norm* linear
Scores (local quick bench, 0-shot chat)
MMLU 200 / ARC-Challenge 150 / GSM8K 60 questions, run through llama-server.
| Model | MMLU | ARC-C | GSM8K | Avg |
|---|---|---|---|---|
| Liquid Grand Cru (this) | 40.0 | 62.0 | 78.3 | 60.1 |
| Best donor (Huihui abliterated) | 57.5 | |||
| LiquidAI/LFM2.5-1.2B-Base | 56.3 | |||
| LiquidAI/LFM2.5-1.2B-Instruct (stock) | 52.1 |
Small-sample numbers from our own harness, so treat them as directional only.
Use
Works in llama.cpp (builds with LFM2 support), Jan, LM Studio and the Winery app.
llama-cli -m Winery-LFM2.5-1.2B-GrandCru-Q8_0.gguf -cnv
Licensed under the LFM Open License v1.0 inherited from LiquidAI. One donor is abliterated, so it refuses less than stock.