Qwen3.5-4B tuning run blows up

This is less a “Qwen is broken” story than a reminder that fresh open models often ship faster than the best fine-tuning playbooks around them.

• –The post is a barebones failure report, so the main value is as an early signal for other developers trying to tune Qwen3.5-4B right now
• –For small reasoning-heavy models, instability can come from setup details like precision, learning rate, masking, packing, or chat-template mismatches rather than the base weights alone
• –Qwen3.5’s recent release cadence means community tooling and training defaults may still be catching up, especially for coding and reasoning datasets
• –With no comments or confirmed fix yet, this reads as a watch item for the open-model ecosystem rather than a verified model-wide defect