Atlas

Models

← All models

Motif-3-Beta

Motif Technologies · 2026-07-20 · 314.8B parameters

Motif-3-Beta is a preview checkpoint for the in-house sparse Mixture-of-Experts causal language model developed by Motif Technologies. The released checkpoint contains 314,841,775,750 parameters; the model card reports about 13 billion active parameters per token and top-8 routing across 384 experts plus one shared expert. Its chat template can enable or suppress a reasoning trace, and it supports multilingual text generation with a native 262,144-token context, Grouped Differential Latent Attention, Grouped PolyNorm, and modified mHC components.

Benchmark scores

BenchmarkScore
AA-LCR61.3
AA-Omniscience Index-17.3
Artificial Analysis Agentic Index34.9
Artificial Analysis Coding Index62.0
Artificial Analysis Intelligence Index44.1
Artificial Analysis Omniscience Accuracy21.0
Artificial Analysis Omniscience Hallucination Rate48.5
Capability154.4
CritPt6.0
GDPval-AA v21259
GDPval-AA v2 (normalized)37.9
GPQA Diamond86.9
Humanity's Last Exam (Text-Only)38.2
SciCode (Artificial Analysis)44.3
Terminal-Bench 2.170.8
𝜏³-Banking28.9
Loading Atlas data…