Summary
Fahd Mirza puts two of China’s newest models — Kimi K3 from Moonshot AI and Qwen 3.8 from Alibaba — through a live head-to-head comparison using real coding tasks rather than synthetic benchmarks. The central test involves a full-stack C port freight tracker application with a deliberately planted bug: a vessel birth clearance check that applies a safety margin backwards, allowing dangerous vessel assignments that should be blocked. Both models run autonomously through the Hermes agent framework, with no hand-holding allowed.
Both Kimi K3 and Qwen 3.8 found and fixed the bug independently. But the comparison reveals sharply different reasoning styles. K3 went beyond the assigned task, reading docstrings, identifying contradictions between comments and frontend behavior, and surfacing three additional real bugs — but looped excessively on confirmation checks, burning 65 API calls and 53 tool calls. Qwen 3.8 stayed narrowly scoped to the one assigned bug, completed it in 45 API calls and 42 tool calls, used roughly half the cached tokens, and exited cleanly on its first successful judge check.
Mirza frames the result not as a winner-loser verdict but as a tradeoff: K3 is more thorough and creatively explores beyond its brief; Qwen 3.8 is leaner, more disciplined, and cheaper to run. The video also notes the broader context of Chinese AI momentum, with DeepSeek V4 GA still approaching general availability. A multimodal comparison round is included alongside the coding task.
📺 Source: Fahd Mirza · Published July 20, 2026
🏷️ Format: Comparison
