2 September 2026

Qwen model trained on 1,928 work tasks, performance improved 70%

• Mercor and SkyRL companies took Qwen3.5-397B-A17B model, a large language model with 397 billion parameters, and trained it on 1,928 different knowledge work tasks like analysis and writing.
• The trained model's APEX-Agents Pass@1 score, a measure of how often it completes tasks correctly on first try, increased by 70 percent.
• The companies found that how they built the training environment, managed computing resources, and designed the system mattered as much as the underlying algorithm itself.

How it was covered

TLDR AITLDR editorial team

Mercor and SkyRL post-trained Qwen3.5-397B-A17B on 1,928 expert knowledge-work tasks, lifting APEX-Agents Pass@1 by 70%. The work shows that robust environments, token accounting, async RL, and harness design matter as much as algorithm choice at frontier scale.