Snowflake lifts a small coding agent from 47.4% to 62.0% in tests

Snowflake AI Research made Arctic-SWE, a method to train coding agents with reinforcement learning. The agent gets a reward when the tests accept its patch. Snowflake reports that it moved the model Qwen3.5-4B from 47.4% to 62.0% on SWE-bench Verified, a test of bug repairs. It also reports that almost a third of training tasks in 3 open data sets accept an incorrect patch.

How it works

Arctic-SWE is a training method that does not wait for other work. Snowflake first tests if the method can overfit 160 tasks. To overfit is to learn a small set of tasks fully and hold the result. The method ends a rollout with no reward at the first incorrect tool use. It also sends work to a free engine, which cut the time for each update from 22 to 11.6 minutes.

The numbers

• Qwen3.5-4B went from 47.4% to 62.0% on SWE-bench Verified, on one eight-GPU server.
• FrogNano, a work at the same time, gets 61.5% with 256,000 rollouts.
• Arctic-SWE got this result after 43,520 rollouts.

What is not known

No other group tested these results. An AI agent that Snowflake made wrote the incorrect patches in the audit. Snowflake gives no data on a large model.

Insight: Weak tests can limit coding agents more than weak models

Snowflake reports that 5,351 of 17,861 approved tasks give a reward for an incorrect patch. One more story shows that good data gives more SWE-bench Verified points. This reads as a change in the limit: the test, not the model, is the key part. Teams that audit their tests can get an edge.