Snowflake releases Arctic-Text2SQL-R2 alongside ZoRRo, a framework that speeds up enterprise SQL reinforcement learning by 3.5x

Original post

I’m excited to share new research work from the Snowflake AI Research Team focused on advancing enterprise AI systems.

Arctic-Text2SQL-R2 is a reasoning model designed for enterprise SQL generation. Trained on Snowflake-native data and optimized for real-world enterprise SQL workloads, the specialized model outperforms larger frontier models on difficult SQL benchmarks despite being 30–150x smaller than other high-performing models.

To make specialized models like Arctic-Text2SQL-R2 practical at scale, the team also introduced ZoRRo (Zero Redundancy Rollouts), a set of optimizations that eliminate redundant computation in long-context RL workflows. ZoRRo accelerated RL training by up to 3.5x, reducing runtime from over five days to only 1.5 days. It also reduced memory consumption enough to support 3.2x longer context windows, enabling more efficient training on complex enterprise reasoning workloads.

Together, this work demonstrates how the next wave of enterprise AI innovation will be driven by both stronger domain-specific models and more efficient training systems. Read more in the blog posts in the comments:

Arctic-Text2SQL-R2: http://www.snowflake.com/blog/engineering/enterprise-text-to-sql-arctic-r2 ZoRRo: http://www.snowflake.com/blog/engineering/zorro-enterprise-rl-training

@yao_zhewei @yuxionghe @samyamrb @jeffra45 @StasBekman

10:23 AM · May 27, 2026 · 655 Views

VIEWS14.5KBOOKMARKS150LIKES172RETWEETS25REPLIES7

Stas Bekman@StasBekman

This is what we have been working on for the last 6 months or so at the AI Snowflake Research:

Zero Redundancy Rollouts (ZoRRo):

https://www.snowflake.com/en/blog/engineering/zorro-enterprise-rl-training/

If you do RL and you want it to be much faster make sure to have a look.

31d14.5K172150

Ashwinee Panda@PandaAshwinee

@StasBekman is this relevant for agentic coding RL? typically the prompt just isn’t that long relative to all the turns that we do

31d4113

Lewis Tunstall@_lewtun

@StasBekman Very cool work Stas!

Turbo nit: it's arguably special relativity (not general) which states massive particles can't travel faster than the speed of light :)

Stas Bekman@StasBekman

This is what we have been working on for the last 6 months or so at the AI Snowflake Research:

Zero Redundancy Rollouts (ZoRRo):

https://www.snowflake.com/en/blog/engineering/zorro-enterprise-rl-training/

If you do RL and you want it to be much faster make sure to have a look.

31d83160

clem 🤗@ClementDelangue

@StasBekman Very cool work! Might be useful to be integrated into TRL?

Stas Bekman@StasBekman

This is what we have been working on for the last 6 months or so at the AI Snowflake Research:

Zero Redundancy Rollouts (ZoRRo):

https://www.snowflake.com/en/blog/engineering/zorro-enterprise-rl-training/

If you do RL and you want it to be much faster make sure to have a look.

31d1.6K40

Jonathan Chang@ChangJonathanC

@StasBekman minimax also did Prefix Tree Merge

31d2523

Strata@ChainZenit

@StasBekman Sounds like another rollout optimization angle, cool.

31d1861

Stas Bekman@StasBekman

@PandaAshwinee Shouldn't matter what type of RL as long as either the prompt is long or instead your rollout n is quite high

you can easily compute the potential savings by feeding 3 numbers into this:

n_rollout x (prompt_len + response_len) / (prompt_len + n_rollout x response_len)

31d2664

Stas Bekman@StasBekman

@_lewtun Thank you for the kind words and the correction, Lewis!

31d213

Stas Bekman@StasBekman

@ChainZenit This is both (1) rollout generation and (2) the training-side optimizations.

31d1241

Jun Keun Yi@yijunkeun

@StasBekman Great to see the learnings being shared, I should try to adopt this too!

31d181

Stas Bekman@StasBekman

@_lewtun It has been fixed.

31d23

Stas Bekman@StasBekman

Indeed, there are multiple other versions out there. Probably at least 2 more.

While we haven't compared them we hope ours is one of the most efficient ones.

Ours can run with any RL-backend, the whole functionality can be transparent to RL since it's folded into the patched model.

31d22

Philip Bontrager@FilipoGiovanni

@StasBekman For “ZoRRo for Inference”, if you use an inference library like vLLM and request n samples per request instead of sending n duplicate requests, don’t you get this same behavior?

31d11

49 Agents IDE - IDE for Agentic Coding@49agents

@PandaAshwinee @StasBekman the prompt doesnt need to be long when the agent has tools to explore the codebase itself. the real constraint is how well the model holds onto what it finds across turns. which is a dataset problem, not a prompt engineering problem

31d9