Liquid AI's LFM2.5 tops mobile benchmarks: 2.32GB memory, 8s latency on iPhone 17 Pro
maximelabonne · x · 2026-09-22
Artificial Analysis, partnering with Liquid AI, benchmarked 39 small models (fitting in 8GB after quantization) running on-device on iPhone 17 Pro and Galaxy S26 Ultra.
- Four of five LFM2.5 models land on the joint Pareto frontier for intelligence, latency, and memory
- LFM2.5-2.6B ties Nanbeige 3B for top intelligence, beating models 3-10x larger with 40% less memory and 3x lower latency
- On iPhone: 2.32GB peak memory, 8.0s end-to-end latency vs Nanbeige's 4.03GB / 21.4s; on Galaxy: 2.45GB / 18.6s vs 4.13GB / 71.4s
- Highest-scoring model under 2.5GB on both devices
Scores average five benchmarks (BFCL subset, IFBench, AA-Omniscience, GPQA Diamond, MATH-500) at 16K context; E2E time covers a 1,024-token prompt plus a 256-token generation. Participants include Nanbeige, Alibaba, Google, TII, IBM, and others.
More from Infra
- SiliconBench: speed, memory and fidelity of nine LLM engines on unified-memory desktops — PennState · 2026-09-22
- Engram retrieval won't replace FFNs, but cutting 40-50% of HBM needs is the real win — bookwormengr · 2026-09-22
- Huawei's Sept 19 event: long-running agents bottleneck on data movement, not FLOPs — krishnan · 2026-09-22
- OnSemi Raises Prices Again as Supply Chain Component Costs Climb 15-20% — BenBajarin · 2026-09-22
- Harvard Analyzes 6.12B AI Requests on Chutes: Agents Use More Tokens, Answer Shorter — markjeffrey · 2026-09-22
- Richard Socher: AI Hard-Takeoff Scenarios Underestimate Physical-World Bottlenecks — RichardSocher · 2026-09-22