Veyra AI

community

AI & ML interests

Building tiny language models for practical local AI. Veyra AI focuses on CPU-friendly inference, function calling, tool use, Python-oriented small models, distillation, RLVR, and lightweight fine-tuning. The goal is to make compact models that are easy to run, inspect, adapt, and use in real workflows without large hardware.

Recent Activity

Jdudeo updated a collection 9 days ago

Veyra2

Jdudeo updated a collection 9 days ago

Veyra2

Jdudeo updated a collection 9 days ago

Veyra2

View all activity

Organization Card

Community About org cards

Welcome to Veyra AI

We build tiny language models for fast local inference. We focus on compact CPU-friendly models that are easy to run and fine-tune. Our models can be found on the Open SLM Leaderboard. We are also a member of the Small Language Model Consortium!

Models:

Model	Release Date	Description
Veyra2-Blueberry-5M-Base	2026-08-27	Second generation Veyra base model with 5M parameters.
Veyra2-Blueberry-5M-Instruct	2026-10-02	Second generation instruction-tuned Veyra model with 5M parameters.
Veyra2-Blueberry-10M-Base	2026-08-27	Second generation Veyra base model with 10M parameters.
Veyra2-Blueberry-10M-Instruct	2026-10-02	Second generation instruction-tuned Veyra model with 10M parameters.
Veyra2-Mango-15M-Base	2026-08-08	Second generation Veyra base model with 15M parameters.
Veyra2-Mango-15M-Instruct	2026-10-02	Second generation instruction-tuned Veyra model with 15M parameters and experimental tool-calling support.
Veyra2-Mango-30M-Base	2026-08-08	Second generation Veyra base model with 30M parameters.
Veyra2-Mango-30M-Instruct	2026-10-02	Second generation instruction-tuned Veyra model with 30M parameters and experimental tool-calling support.
Veyra2-Apricot-50M-Base	2026-06-14	Second generation Veyra base model with 50M parameters.

Note on old models: Only the current family of models will been shown on the table, models like Veyra 30M can still be found here.

Datasets:

Dataset	Release Date	Description
Micro-Model-Bench	2026-08-27	A collection of benchmark results for models under 150M.
Sci-Cloze-900	2026-06-14	Cloze-style GCSE Combined Science benchmark for evaluating small base language models.