LFM2-350M: benchmark results
Liquid AI's LFM2-350M, the smallest of four LFM2 checkpoints from 350M to 2.6B parameters, built with a hybrid architecture of multiplicative gates and short convolutions for edge deployment. Provider: Liquid AI. Released 2025-07-10. Access: Open.
Aggregate Rating 1181 ± 18, rank #1608 of 1661 rated models, from 358 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Polish Summarization - PSC | 23.53 | chrF3++ (%) | 76.9 |
| EuroEval Spanish Summarization - Mlsum ES | 26.71 | chrF3++ (%) | 68 |
| Enkrypt AI - Bias Risk | 80.62 | Risk Score | 60.3 |
| EuroEval Portuguese Summarization - Publico | 31.84 | chrF3++ (%) | 59.7 |
| Enkrypt AI - Toxicity Risk | 2.88 | Risk Score | 56.6 |
| EuroEval Albanian Summarization - LR SUM SQ | 28.19 | chrF3++ (%) | 48.4 |
| Kurakurai French LLM - GPQA Diamond | 25.76 | Accuracy (%) | 46.3 |
| EuroEval Serbian Summarization - LR SUM SR | 24.4 | chrF3++ (%) | 43.5 |
| EuroEval Norwegian Knowledge - Idioms NO | 1.78 | MCC (x100) | 42.5 |
| EuroEval German Summarization - Mlsum DE | 30.86 | chrF3++ (%) | 40.5 |
| EuroEval Icelandic Common Sense Reasoning | 3.96 | Common Sense Reasoning Average Score (%) | 39.3 |
| EuroEval Italian Summarization - Ilpost SUM | 33.68 | chrF3++ (%) | 36.7 |
Interactive version: theaggregate.ai/model?slug=lfm2-350m · How It Works · Data refreshed daily, snapshot 2026-10-23.