Deep reasoning
A 256K-token window and step-by-step reasoning that held its logic across long, messy problems most models drop halfway through.
Aurora is a frontier large language model, built in Canada. It reasoned across 256K tokens of context, wrote production code, and served millions of people at frontier quality. In 2026, Aurora became Qai.
aurora-2 / 380B parameters / 256K context / text, vision, audio
Aurora was a general model with a serious head for hard problems. Long context, real code, and a genuine grasp of images and audio, all in one conversation.
A 256K-token window and step-by-step reasoning that held its logic across long, messy problems most models drop halfway through.
Wrote, debugged, and reviewed code across more than 80 languages. It scored 90.1% on HumanEval and closed real GitHub issues.
Read images, charts, screenshots, and audio in the same thread as text. One model, three modalities, no handoffs.
Reasoned and translated fluently across more than 100 languages, holding tone and nuance, not just literal words.
Aurora-2 posted frontier scores across reasoning, code, and math, edging out GPT-4o and Gemini 1.5 Pro on the reasoning benchmarks. Here are four of them. The full head-to-head lives on the benchmarks page.
Everything Aurora built now powers Qai. The models, the research, one place to use them. If you came here to try Aurora, this is where you go next.