Apriel-1.6-15B-Thinker

Name: Apriel-1.6-15B-Thinker
Author: ServiceNow AI Research

ServiceNow AI Research · 2025-12-09 · 14.9B parameters

Apriel-1.6-15B-Thinker is ServiceNow AI Research's December 2025 update to the Pixtral-derived Apriel-1.5 multimodal reasoning checkpoint. It improves text and image reasoning while reducing reasoning-token usage by more than 30% relative to V1.5, supports tool use, and remains deployable on a single GPU. The checkpoint is released under MIT.

Benchmark scores

Benchmark	Score
AA-LCR	50.3
AA-Omniscience Index	-59.4
AIME 2025 (Artificial Analysis avg@10)	88.0
Artificial Analysis Omniscience Accuracy	17.0
Artificial Analysis Omniscience Hallucination Rate	91.9
Artificial Analysis Openness Index	50.0
Artificial Analysis Progress Index v1	234
Capability	140.8
CritPt (avg@5)	0.3
GPQA Diamond (Artificial Analysis avg@5)	73.3
Humanity's Last Exam (Text-Only)	10.8
IFBench (Artificial Analysis avg@5)	69.1
LiveCodeBench (Artificial Analysis)	80.7
MMLU Pro	79.0
SciCode (Artificial Analysis avg@3)	37.3
Terminal-Bench Hard	14.4
τ²-bench (Telecom)	69.3