Snowflake releases Arctic RL with ZoRRo: Text2SQL-R2 training drops to ~36 hours

• Snowflake open-sourced Arctic RL, and StasBekman's launch thread says its ZoRRo optimization can be turned on in VeRL and SkyRL with one config flag, with no code changes.
• According to the launch thread, ZoRRo delivers up to 6x faster actor updates and a 3.5x end-to-end training speedup, cutting Arctic-Text2SQL-R2 training from about five days to about 36 hours on 32 H200 GPUs.
• In the same post, StasBekman said Arctic-Text2SQL-R2 scored 48.7 on Snowflake's evaluated enterprise SQL benchmark, ahead of Gemini 3.1 Pro at 47.9 and Claude 4.7 at 47.3 under Snowflake's tested conditions.
• The release thread also ships two open recipes: a text-to-SQL setup that moved BIRD dev accuracy from 59.92% to 70.35%, and a multi-hop QA recipe that raised average accuracy from 69.6% to 72.3%.

You can jump straight to the Arctic RL project and the separate fast RL framework repo. Snowflake's launch thread includes the training-speed chart, while the companion repo post makes clear that SkyRL and VeRL are only the first integrations.

The headline number here is training throughput. Snowflake said ZoRRo speeds up actor updates by as much as 6x and improves end-to-end RL training by 3.5x, with its own chart showing Arctic RL far below the baseline iteration time.StasBekman's launch thread

That is how Snowflake gets to the flashy example in the post: Arctic-Text2SQL-R2 drops from roughly five days of training to roughly 36 hours on 32 H200s, according to the release thread.

Snowflake split the release into two pieces:

• Arctic RL integrates with VeRL and SkyRL today, per the launch thread.
• ZoRRo can be enabled with one config flag and no code changes, according to StasBekman.
• A separate repo packages the fast RL framework and integration code, which the companion post says is meant to expand to more frameworks.
• The text-to-SQL recipe improved BIRD dev accuracy from 59.92% to 70.35%, per the release thread.
• The multi-hop QA recipe improved average accuracy from 69.6% to 72.3%, again from Snowflake's launch post.

Snowflake also used the release to make a model-quality claim: the same thread puts Arctic-Text2SQL-R2 at 48.7 on Snowflake's enterprise SQL benchmark, versus 47.9 for Gemini 3.1 Pro and 47.3 for Claude 4.7 under the tested setup.

The project was not a quick repo dump. Replying to a question about development time, StasBekman said the work took "probably about 6+ months."

Snowflake also appears to have more formal writeup coming. In another reply, StasBekman said the plan is to publish a paper once the person leading that effort is back from vacation.