LFM2 8B A1B: benchmark results
Liquid AI's LFM2-8B-A1B, their first MoE model in the LFM2 hybrid edge-deployment family, with 8.3B total and 1.5B active parameters, meant for fine-tuning on narrow agentic or RAG tasks. Provider: Liquid AI. Released 2025-10-07. Access: Open.
Aggregate Rating 1387 ± 19, rank #1263 of 1656 rated models, from 170 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| QIMMA - ArabLegalQA | 72 | BERTScore F1 (%) | 89 |
| MERA Code - CodeCorrectness | 82.59 | Accuracy (%) | 81.5 |
| QIMMA - MedArabiQ QA | 53.2 | BERTScore F1 (%) | 74.6 |
| QIMMA - 3LM MBPP+ | 59.52 | pass@1 (%) | 67.1 |
| QIMMA - 3LM HumanEval+ | 61.59 | pass@1 (%) | 64 |
| QIMMA - ArabCulture | 49.87 | Accuracy (%, log-likelihood multiple choice) | 62.2 |
| MERA Code - stRuCom | 24.8 | chrF (%) | 58.5 |
| QIMMA - MedArabiQ MCQ | 36.84 | Accuracy (%, log-likelihood multiple choice) | 54.9 |
| QIMMA - MizanQA | 53.76 | Normalized multiple-choice probability score (%) | 51.2 |
| UGI - Willingness (W/10) | 5.8 | W/10 Score | 50.8 |
| MERA - LCS | 12.6 | Accuracy (%) | 48.6 |
| QIMMA - MedAraBench | 35.02 | Accuracy (%, log-likelihood multiple choice) | 47.6 |
Interactive version: theaggregate.ai/model?slug=lfm2-8b-a1b · How It Works · Data refreshed daily, snapshot 2026-10-20.