Harmonic-9B drops with two-stage Qwen3.5 reasoning, agentic focus

Harmonic-9B proves that high-signal data curation, not just scale, is the secret to making small models behave like reliable agents.

• –Two-stage training strategy (Reasoning Foundation then Agentic Specialization) ensures the model understands "why" it calls tools, not just "how."
• –The filtered Hermes dataset increases self-correction markers from 6% to 63%, addressing the most common failure point in long-running agent loops.
• –Thinking depth is enhanced by 40%, moving beyond shallow single-hop tool usage to multi-path exploration and logic verification.
• –Structural quality pipeline ensures 100% valid JSON for tool calls, a critical stability requirement for local agent deployments.
• –Optimized for the Qwen3.5-9B architecture, providing a competitive reasoning backbone that fits on consumer-grade hardware.