Gemma 4 coding: Harness choice swings performance

Gemma 4 marks the end of "vibe coding" benchmarks as the agentic harness becomes just as important as the model weights.

• –The 31B Dense model hits 80% on LiveCodeBench v6, rivaling proprietary frontier models in a local-first package.
• –Frameworks like Kilo Code and Roo Code extract more performance through highly structured system prompts and autonomous tool execution.
• –Variation in scores across harnesses (Claude Code vs. Kilo Code) suggests that "raw" model evals are increasingly decoupled from real-world agentic utility.
• –26B MoE variant is the sweet spot for developers, offering 97% of Dense performance at a fraction of the inference cost.