Atlas

Models

← All models

LFM2.5-1.2B-Thinking

Liquid AI · 2026-01-20 · 1.2B parameters

LFM2.5-1.2B-Thinking is Liquid AI’s dense 1,170,340,608-parameter text-only reasoning checkpoint, released on January 20, 2026 for on-device deployment. Derived from LFM2.5-1.2B-Base, its training recipe uses reasoning-trace mid-training, supervised fine-tuning on synthetic traces, preference alignment, and RLVR; at inference it emits thinking traces before answers, with Liquid reporting its largest gains over the Instruct checkpoint on math, instruction following, and tool use. Its 16-block hybrid backbone contains 10 double-gated short-convolution blocks and six GQA blocks, and Liquid documents a 32,768-token context and phone deployments around 0.9 GB depending on runtime and quantization.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-83.9
Artificial Analysis Intelligence Index2.7
Artificial Analysis Omniscience Accuracy6.6
Artificial Analysis Omniscience Hallucination Rate96.9
Artificial Analysis Openness Index27.8
Capability114.6
CritPt0.0
GPQA Diamond33.9
Humanity's Last Exam (Text-Only)6.1
IFBench (Artificial Analysis)41.8
SciCode (Artificial Analysis)4.2
Terminal-Bench Hard0.0
τ²-bench (Telecom)19.6
Loading Atlas data…