The Qwen3.8 27B model demonstrated performance comparable to advanced AI models in SQL benchmarks while running locally on hardware with limited resources. Testing was conducted using the DABstep benchmark, which evaluates how models handle data-related queries and business context.
Running Qwen3.8 27B locally costs only the price of electricity, whereas using the GPT 5.6 Luna Max model for similar tasks costs over $8 per run. The local setup allows for significant cost savings, with the author noting a difference of more than 17 times in expenses compared to cloud-based alternatives.
For local execution, the author utilized a MacBook Air with an M5 chip and 32 GB of RAM for the 4-bit version of the model, and a MacBook Pro with an M1 Pro chip and 16 GB of RAM for the 3-bit version. The 4-bit version achieved 98.6% accuracy, while the 3-bit version reached 96.4%.
Performance metrics showed that the local model generates approximately 5–7 tokens per second. While cloud models like Gemini-3-Flash or GPT 5.6 Luna respond faster—taking 25 to 40 seconds compared to 5–6 minutes for the local model—the local approach eliminates the risk of unexpected high costs.
Hardware requirements for this setup include at least 16 GB of unified memory or a discrete GPU with 16 GB of VRAM. The author suggests that 32 GB of RAM provides a noticeable performance advantage, and notes that Apple Silicon devices are particularly effective due to high memory bandwidth.
