Poolside Laguna S 2.1 is worse than Qwen 3.6 27B and Gemma4 31B

Article automatically generated from technical news.

I did head to head comparison between Laguna S 2.1, Qwen 3.6 27B and Gemma4 31B. Setup: Qwen 3.6 27B (fp8) and Gemma4 31B q6 on my RTX PRO 5000, Laguna S2.1 q4/5/6 on a single DGX Spark. Speed was fine: NVFP4 ~27 tok/s (peaks ~39), Q5/Q6 ~14 tok/s. The Q6 really pushed the spark with 124GB mem in use, but it didn't crash. Ran all of them through the same task local bench - HTML/canvas mini apps, tool calling, Python, prose and each output scored blind (models anonymised, r