Atlas

Models

← All models

LFM2-2.6B

Liquid AI · 2025-09-23 · 2.6B parameters

LFM2-2.6B is Liquid AI's 2,569,272,320-parameter dense, post-trained LFM2 checkpoint for on-device chat, reasoning, and tool use. Its 30-layer hybrid backbone interleaves 22 gated short-convolution blocks with eight grouped-query-attention blocks, and the model can emit dynamic `<think>`/`</think>` traces for complex or multilingual prompts. Liquid's model card and technical report specify a 32,768-token supported context, although the current Hugging Face config sets `max_position_embeddings` to 128,000.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-52.1
AIME 20258.3
Artificial Analysis Intelligence Index2.7
Artificial Analysis Omniscience Accuracy5.2
Artificial Analysis Omniscience Hallucination Rate60.5
Artificial Analysis Openness Index33.3
Capability101.3
CritPt0.0
Global MMLU Lite34.3
GPQA Diamond30.6
Humanity's Last Exam (Text-Only)5.2
IFBench (Artificial Analysis)19.5
LiveCodeBench (Artificial Analysis)8.1
SciCode (Artificial Analysis)2.5
SnakeBench Average Apples0.5
SnakeBench Best Apples2.0
SnakeBench Rating13.4
SnakeBench Total Apples7.0
SnakeBench Win Rate41.7
Terminal-Bench Hard0.8
τ²-bench (Telecom)13.5
Loading Atlas data…