Atlas

Models

← All models

LFM2.5-1.2B-Instruct

Liquid AI · 2026-01-05 · 1.2B parameters

LFM2.5-1.2B-Instruct is Liquid AI's general-purpose, text-only instruction-tuned checkpoint for on-device and edge deployment. It extends the LFM2 pretraining budget from 10T to 28T tokens, then applies supervised fine-tuning, preference alignment, and large-scale multi-stage reinforcement learning for instruction following and tool use. Its 16-layer hybrid backbone contains 10 double-gated LIV convolution blocks and six GQA blocks; although the released config sets max_position_embeddings to 128,000, the official checkpoint card documents a 32,768-token context length.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-73.8
Artificial Analysis Intelligence Index2.7
Artificial Analysis Omniscience Accuracy6.0
Artificial Analysis Omniscience Hallucination Rate84.8
Artificial Analysis Openness Index27.8
BuseyBench SVG0.4
Capability110.9
CritPt0.0
GPQA Diamond32.6
Humanity's Last Exam (Text-Only)6.8
IFBench (Artificial Analysis)43.8
SciCode (Artificial Analysis)2.3
Terminal-Bench Hard0.0
τ²-bench (Telecom)10.8
Loading Atlas data…