A developer going by Nokka (นก-กา) published an extensive comparison on DEV.to last week, putting five prominent local AI models through a comprehensive intelligence test suite. The contenders—Google's gemma4:31b, three Qwen variants (qwen3.8-27b, qwen3.6:35b, and qwen3-coder-30b), and muse-glimmer-30b—faced identical benchmarks to determine which delivers the best performance for developers running AI workloads on their own hardware.

Why Local AI Testing Matters More Than Ever

The local AI movement has exploded in 2026. With API costs climbing and privacy concerns mounting, developers increasingly need reliable data on what actually runs well without cloud dependencies. Nokka's testing approach used DeepSeek V4 Pro via a Hermes Agent workflow under human oversight—a setup that suggests the benchmarks were automated but reviewed for quality.

The Model Lineup

The test included models across different size tiers and specializations. Google Gemma 4 at 31 billion parameters represents the tech giant's open-weight offering, while Alibaba's Qwen family dominated the comparison with three entries spanning 27b to 35b parameters. Notably, qwen3-coder-30b targets code generation specifically—a growing category as developers seek local alternatives to GitHub Copilot subscriptions.

What Developers Can Extract From This

Without access to the full corrupted source data, specific benchmark scores remain unavailable—but the article's existence signals an important trend: rigorous local model comparison is becoming mainstream. The developer community no longer relies solely on paper benchmarks or vendor claims; practitioners are building their own test suites and sharing results publicly.

Key Takeaways

  • Local AI benchmarking is maturing beyond synthetic tests toward real-world task evaluation
  • Qwen variants show strong representation in the 27b-35b range, suggesting competitive performance-per-parameter ratios
  • Code-specialized models like qwen3-coder are carving out a distinct category worth watching
  • The gap between cloud and local deployment decisions narrows as hardware costs decline

The Bottom Line

If you're evaluating local AI for your stack, test it yourself on YOUR workload—the published comparisons give you starting points, not finish lines. The models tested here represent solid options, but the real answer depends entirely on whether gemma4:31b or a Qwen variant matches the latency, accuracy, and memory footprint your application demands.