Back to models
Foundation · Pre-trained2026Foundation model

AQ-5B

A foundation model by Zyora Labs

AQ-5B is Zyora Labs' own foundation model — a decoder-only Transformer trained from scratch in pure PyTorch to understand every educational stream, from Physics and Maths to Botany, Tamil and Engineering. It is the base that will power the next generation of ZyoraByte and Zyora AI.

5B
Parameters
32K
BPE tokenizer vocabulary
EN · TA
Bilingual — English & Tamil
Scratch
Pre-trained from zero
Model card
Type
Foundation model — decoder-only, pre-trained from scratch
Parameters
5 billion
Architecture
Decoder-only Transformer — RoPE + RMSNorm + SwiGLU + GQA
Implementation
Pure PyTorch (hand-coded, not HuggingFace transformers)
Tokenizer
Custom BPE · 32,000 vocabulary
Languages
English & Tamil
Domain
All educational streams (base → advanced concepts)
Developer
Zyora Labs
Team
V
Vasanth
Chief Researcher
AS
Adithi S
Junior AI Engineer
Overview

What is AQ-5B?

AQ-5B is Zyora Labs' own foundation model — built from scratch, not fine-tuned from someone else's weights. It is a decoder-only Transformer designed to understand every educational stream, from Physics and Maths to Botany, Zoology, Tamil, Engineering and Arts & Science.

Rather than memorising a specific syllabus, AQ-5B is trained on broad conceptual knowledge — base to advanced, so it forms a durable foundation for the next generation of ZyoraByte and Zyora AI products.

Architecture

A modern decoder, hand-built

AQ-5B is a decoder-only Transformer RoPE + RMSNorm + SwiGLU + GQA — implemented in pure PyTorch, hand-coded in torch.nn rather than pulled from HuggingFace transformers. The config is kept HF-compatible so the model can be loaded for inference and sharing later.

  • Rotary position embeddings (RoPE) for long-range context.
  • RMSNorm for stable, efficient normalisation.
  • SwiGLU feed-forward blocks.
  • Grouped-query attention (GQA) for efficient inference.
  • Custom BPE tokenizer with a 32,000-token vocabulary.
Coverage

Bilingual, every stream

AQ-5B is bilingual in English and Tamil and treats all educational streams as first-class — the sciences, mathematics, languages, engineering and the arts. It is trained on conceptual knowledge from foundational to advanced, giving downstream tutors a broad, reliable base to reason from.

Responsible use

Scope

AQ-5B is a foundation model — a base intended for further training and alignment rather than a finished consumer product. It is built to give downstream tutors like ZyoraByte a broad, reliable base to reason from across every educational stream.

FAQ

Frequently asked questions

What is AQ-5B?

AQ-5B is Zyora Labs' own foundation model — a decoder-only Transformer trained from scratch to understand every educational stream. It is the base model for the next generation of ZyoraByte and Zyora AI products.

What architecture does AQ-5B use?

AQ-5B is a decoder-only Transformer using RoPE, RMSNorm, SwiGLU and grouped-query attention (GQA), hand-coded in pure PyTorch rather than HuggingFace transformers.

How large is AQ-5B?

AQ-5B is a 5-billion-parameter foundation model built from scratch by Zyora Labs.

Which languages and subjects does AQ-5B cover?

AQ-5B is bilingual in English and Tamil and is trained on broad conceptual knowledge across all educational streams — Physics, Chemistry, Botany, Zoology, Maths, English, Tamil, Engineering, Arts and Science.

Who is building AQ-5B?

AQ-5B is built by Zyora Labs, led by Vasanth (Chief Researcher) with Adithi S (Junior AI Engineer).

Want to build with AQ-5B?