Renaissance Labs

Atlas / North America

Academic lab · Berkeley, United States

NovaSkyBerkeley Sky Lab

Berkeley Sky Computing Lab group releasing open reasoning models, training recipes and the SkyRL reinforcement-learning library for LLM agents.

reasoningreinforcement learningdatasets

Berkeley, United States

Founded
2025
Type
Academic lab
Based in
Berkeley, United States
Key papers
2
Featured projects
6

About

NovaSky is a team within the Sky Computing Lab at UC Berkeley. Its GitHub organization was created in January 2025, the month it published Sky-T1, a reasoning model it said could be trained for under 450 dollars.

The group releases weights, data and training code together. After Sky-T1 it published work on cutting reasoning costs, test-time scaling for code generation and the SkyRL framework, which grew from SkyRL-v0 for long-horizon software agents into a modular RL library for training LLMs.

Its repositories, including SkyThought and SkyRL, are released under Apache 2.0, and its audience is researchers and developers training reasoning and agentic models.

Mission and ambitions

NovaSky aims to make post-training of reasoning and agentic models reproducible and affordable by sharing models, data and code openly.

Known for

  1. 01

    Sky-T1 reasoning model trained for under 450 dollars

  2. 02

    SkyRL reinforcement-learning library

  3. 03

    SkyThought open recipes

  4. 04

    Text-to-SQL and software-agent RL results

Milestones

  1. 2025

    Sky-T1-32B-Preview

    An open 32-billion-parameter reasoning model, with data and code, reported to cost under 450 dollars to train.

  2. 2025

    Reasoning cost reduction

    A January 2025 post showed how to cut reasoning costs by about half without sacrificing accuracy.

  3. 2025

    SkyRL-v0

    A May 2025 release for training long-horizon real-world agents with reinforcement learning.

  4. 2025

    SkyRL-SQL

    A May 2025 recipe for data-efficient multi-turn reinforcement learning on text-to-SQL.

  5. 2025

    SkyRL-v0.1

    A June 2025 release evolving SkyRL into a modular, performant RL framework for LLMs.

  6. 2025

    SA-SWE-32B

    A 32-billion-parameter software-engineering model published in October 2025.

Open source

GitHubgithub.com/NovaSky-AI
Hugging Facehuggingface.co/NovaSky-AI

Research papers

Notable releases

Openness

Open weights, data and code.

People

Listed from the lab's own pages and public profiles.

Contact and community

Contact details and links are on the lab's website (https://novasky-ai.github.io); code questions go through issues on their GitHub organisation.

Public Discord linked from the lab's own website for community discussion and announcements.

Sources

Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.

Labs nearby