Qwen 3.6 benchmarks, Gemma 4 wins practice

Benchmaxing looks real here: the writeup argues that leaderboard wins do not necessarily translate into better developer experience, especially once you leave curated evals and start feeding models ugly inputs. The most useful signal is not raw accuracy, but how well a model survives instruction-following, token efficiency, and video/OCR pipeline friction.

• –Qwen 3.6 reportedly overthinks less on easy prompts, but still spirals on obscure visual tasks and can waste thousands of tokens before failing to answer.
• –Gemma 4 is presented as stronger at structured outputs, bounding boxes, and normalized coordinate handling, which matters more than raw benchmark scores for production pipelines.
• –The post suggests a regional-data split: Gemma handles Western/European references better, while Qwen is stronger on Asian context and some GeoGuessr-style identification.
• –For video, Qwen gets credit for better exercise tracking and rep counting, but both models remain shaky on deepfake-style classification.
• –The vLLM inference takeaway is practical: visual token budget settings can make Gemma look much worse than it is, so default engine knobs matter as much as model choice.