Atlas

Models

← All models

Tiny Aya Global

Cohere · 2026-02-17 · 3.3B parameters

Tiny Aya Global is Cohere Labs' globally balanced instruction-tuned Tiny Aya checkpoint, post-trained from a dense 3.35B-parameter base pretrained for 6 trillion tokens; Cohere describes the family as covering 70 languages, although this checkpoint's model-card metadata enumerates 67 language tags. It uses a 36-layer decoder-only Transformer with 3:1 interleaved 4,096-token sliding-window RoPE layers and global NoPE layers plus grouped-query attention. Cohere targets local and on-device multilingual generation; its documentation and technical report support 8,192-token input and output contexts, although the official GGUF conversion declares 500,000 maximum positions.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-84.7
Artificial Analysis Intelligence Index1.0
Artificial Analysis Omniscience Accuracy5.7
Artificial Analysis Omniscience Hallucination Rate95.8
Artificial Analysis Openness Index36.1
Capability105.3
CritPt0.0
Global MMLU Lite55.1
Global PIQA - Non-Parallel (Strict Exact Match)59.5
Global PIQA - Parallel (Strict Exact Match)33.7
GPQA Diamond30.5
Humanity's Last Exam (Text-Only)5.2
IFBench (Artificial Analysis)20.1
SciCode (Artificial Analysis)3.6
Terminal-Bench Hard0.0
τ²-bench (Telecom)0.0
Loading Atlas data…