ServiceNow-AI/SuperApriel-15B-Instruct · Hugging Face

ServiceNow released SuperApriel-15B-Instruct, a single-checkpoint 15B model with 8 deployment presets spanning 1.0× to 10.7× decode throughput at 32K sequence length. It has 48 decoder layers with 4 mixer variants per layer and up to 262K context positions depending on runtime; the key point is that speed-quality tradeoffs and speculative decoding are exposed from the same weights.

Why it matters: A single checkpoint spanning 8 deployment presets with 1.0x-10.7x decode throughput gives strong HKR-H and HKR-K, and the serving tradeoff gives HKR-R. The blast radius is narrower: this is a 15B inference-focused release, not a frontier-lab flagship update, so 76 and featured.