Benchmarks· Technical Case Study

Benchmarking Sub-30B Coding Models on Real-World Refactoring Tasks

A rigorous evaluation of small open-weight coding models on multi-file code navigation, AST manipulation, and edge-case test generation.

Why Sub-30B Models Matter

While frontier models like Claude 3.7 Sonnet and GPT-4o lead in general intelligence, running local sub-30B parameters models on developer workstations provides significant privacy, latency, and cost benefits for continuous code completion and localized refactoring.


Benchmark Methodology

We evaluated models across 100 complex refactoring scenarios extracted from production TypeScript and Rust codebases:

| Metric | Target | Hermes-3 (8B) | Qwen-2.5-Coder (14B) | DeepSeek-Coder (33B) | | :--- | :--- | :--- | :--- | :--- | | Syntax Validity | 100% | 94.2% | 98.6% | 99.1% | | Multi-File Context Retention | $ge 90%$ | 78.4% | 91.2% | 93.8% | | Pass@1 on Unit Tests | $ge 80%$ | 68.5% | 82.4% | 85.0% |


Findings & Recommendations

For local IDE integrations where inference speed is paramount, Qwen 2.5 Coder 14B offers the best balance of context comprehension, low memory footprint, and high syntax reliability.

Indexed Topics:Local LLMsBenchmarksEvaluationOpen Source