Ornith-1.5 pushes self-improving coding models

Ornith-1.5’s most interesting contribution is moving self-improvement from a slogan toward a concrete training pipeline, though real-world performance still needs broader independent validation.

• –Self-generated tasks could reduce reliance on expensive human-curated coding datasets
• –Task-specific scaffolds let the model adapt planning, tool use, and recovery strategies to different coding environments
• –The 9B model improves local deployment prospects, while MoE variants target higher-end agentic coding workloads
• –Early community testing suggests promising local performance, but benchmark claims should be treated cautiously until independently replicated