Atlas

Models

← All models

LFM2-8B-A1B

Liquid AI · 2025-10-07 · 8.3B parameters

LFM2-8B-A1B is Liquid AI's early sparse Mixture-of-Experts checkpoint with 8.3 billion reported total parameters and 1.5 billion reported active parameters per token, designed for on-device deployment and tool use. It uses the LFM2 hybrid backbone with 18 gated short-convolution blocks and six GQA blocks; all but the first two layers replace dense feed-forwards with 32-expert, top-4 SwiGLU MoE blocks routed by normalized-sigmoid gating with adaptive biases. Liquid AI reports a 12T-token initial pre-training stage followed by 1T-token long-context mid-training at a 32,768-token context length.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-75.1
AIME 202525.3
Artificial Analysis Intelligence Index1.8
Artificial Analysis Omniscience Accuracy6.8
Artificial Analysis Omniscience Hallucination Rate88.0
Artificial Analysis Openness Index33.3
Capability113.3
CritPt0.0
GPQA Diamond34.4
Humanity's Last Exam (Text-Only)4.9
IFBench (Artificial Analysis)26.3
LiveCodeBench (Artificial Analysis)15.1
MMLU Pro50.5
SciCode (Artificial Analysis)6.8
SnakeBench Average Apples0.5
SnakeBench Best Apples4.0
SnakeBench Rating15.1
SnakeBench Total Apples11.0
SnakeBench Win Rate35.0
Terminal-Bench Hard0.0
τ²-bench (Telecom)10.5
Loading Atlas data…