MiniCPM5-2B: benchmark results

A dense LlamaForCausalLM checkpoint with 2,516,756,480 total parameters, including 1,981,982,720 non-embedding parameters, and native 131,072-token context. This is the final post-trained release rather than a base or SFT-only checkpoint. Provider: OpenBMB. Released 2026-09-07. Access: Open.

Aggregate Rating 1558 ± 20, rank #1458 of 10524 rated models, from 85 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
LFM3-1.7B Release - τ³-bench (4 domains), average69.44Score (%)100
AA-Omniscience Hallucination Rate21.91Hallucination Rate (%)97.8
LFM3-1.7B Release - AA-Omniscience (public)-5Omniscience net score80
LFM3-1.7B Release - AA-Omniscience non-hallucination (public)84.44Score (%)80
LFM3-1.7B Release - IFBench62.08Score (%)80
AA Omniscience-11.65Score62
Docbench - Planted Problem Recall - Substantial13Planted problems found (%)61.1
LFM3-1.7B Release - BFCL V462.65Score (%)60
LFM3-1.7B Release - GPQA Diamond65.15Score (%)60
LFM3-1.7B Release - Humanity's Last Exam (no tools)8.2Score (%)60
LFM3-1.7B Release - LiveClaw30.37Score (%)60
LFM3-1.7B Release - LiveCodeBench v670.52Score (%)60

Interactive version: theaggregate.ai/model?slug=minicpm5-2b · How It Works · Data refreshed daily, snapshot 2026-11-22.