Atlas

Models

← All models

LFM2.5-350M

Liquid AI · 2026-03-31 · 354M parameters

LFM2.5-350M is Liquid AI's March 31, 2026 successor to LFM2-350M, keeping its 16-layer hybrid of 10 gated short-convolution and six grouped-query attention blocks while extending pre-training from 10T to 28T tokens and adding large-scale multi-stage reinforcement learning. The card gives a 32,768-token context and nine languages. The BF16 checkpoint contains 354,483,968 parameters, and the tokenizer assigns 64,402 distinct IDs within 65,536 embedding rows.

Benchmark scores

BenchmarkScore
BFCL v3 Overall Accuracy44.1
BFCL v4 Overall Accuracy21.9
Capability94.0
CaseReportBench - release-reported, metric unspecified32.5
GPQA Diamond30.6
IFBench - release-reported instruction following40.7
IFEval - release-reported, metric unspecified77.0
MMLU Pro20.0
Multi-IF44.9
τ²-bench (Retail)17.8
τ²-bench (Telecom)18.9
Loading Atlas data…