Apriel 15B Thinker

ServiceNow's cost-efficiency flagship: a 15B multimodal (text+image) reasoning model built by depth-upscaling Mistral's Pixtral-12B (40 → 48 layers) — no pretraining from scratch — then staged continual pretraining and text-only SFT with explicit reasoning traces. The 1.5 tech report's tagline: "Mid training is all you need!" — no RL at all in 1.5, trained on 640 H100s in 7 days, fits on a single GPU. MIT license.

Apriel-1.6 (December 9, 2025) expands the CPT mixture, extends text CPT to 49K sequence length, grows SFT to 2.4M samples, and adds multi-stage RL with verifiable rewards (GSPO) — cutting reasoning-token usage by 30%+. Release-era claims: AA index 52 (1.5) → 57 (1.6) on AA's pre-recalibration v3.0 — "at least 1/10 the size of any other model" above 50, and the leading <40B open-weights model. On the recalibrated AA v4.1 both checkpoints read ~21. AIME25: 88, GPQA-Diamond: 73, MMLU-Pro: 79 (1.6, self-reported).