Capabilities

What Aurora
could do.

Aurora is a general-purpose frontier model. It reasoned, coded, saw, listened, and spoke more than 100 languages, all from one set of weights. Here is the full picture. To put it to work today, you use Qai.

01

Reasoning that holds the thread

Aurora carried a 256K-token context, roughly 500 pages, and reasoned across all of it at once. It planned before it answered, showed its work when asked, and caught its own mistakes mid-solution. On graduate-level science questions (GPQA Diamond) it scored 60.2%, ahead of GPT-4o and Claude 3.5 Sonnet.

256K context Step-by-step planning 60.2% GPQA Diamond
02

Code across whole repositories

Aurora wrote, refactored, and reviewed code in more than 80 languages. It scored 90.1% on HumanEval and resolved 51.4% of real issues on SWE-bench Verified, editing across whole repositories instead of one file at a time.

80+ languages 90.1% HumanEval 51.4% SWE-bench Verified
03

Sees, reads, and hears

Text, images, and audio in one model. Aurora read charts, screenshots, handwriting, and diagrams, transcribed and reasoned over audio, and answered about all of it in a single conversation. No separate vision model, no handoffs.

Vision Audio Documents and charts
04

Fluent in 100+ languages

Aurora reasoned and translated across more than 100 languages, holding tone and nuance rather than swapping words one for one. On multilingual grade-school math (MGSM) it scored 86.9%, close to its English performance.

100+ languages 86.9% MGSM Tone-aware translation
05

Tools, functions, and agents

Aurora called tools, APIs, and functions on its own, chained multi-step tasks, and knew when to stop and ask. It ran as the brain of autonomous agents that searched, booked, and wrote back to real systems without a human in every loop.

Function calling Multi-step planning Agent runtime
06

Safe under pressure

Aurora Guardrails filtered harmful output, resisted jailbreaks, and kept the model steerable when prompts got adversarial. Every release went through red-team review before it shipped to a single user.

Jailbreak-resistant Red-teamed Steerable
Model card

Aurora-2, at a glance.

ArchitectureSparse Mixture-of-Experts, 128 experts, 8 active per token
Parameters380B total, 41B active per token
Context window256K tokens, up to 1M on Aurora-2 Reason
ModalitiesText, vision, audio
Training data14.2T tokens, filtered and deduplicated
VariantsAurora-2 Flash, Aurora-2, Aurora-2 Reason
Serving speed320 tokens per second on Flash, 280ms median first token
AvailabilityNow delivered through Qai at q-ai.ca
Put it to work

These capabilities
ship as Qai.

Aurora is not sold on its own anymore. The models, the multimodal stack, and the agent runtime are all delivered through Qai.