Gemma4 vs Qwen3.5 and Qwen3.6 Kebab Test

On April 22nd, someone in the LocalLLaMA community ran what the open-weights crowd calls a Kebab Test — a deceptively simple-sounding benchmark that has quietly become one of the more honest stress tests for instruction-following models. The matchup: Google's Gemma 4 against Alibaba's Qwen 3.5 and the barely-surfaced Qwen 3.6. The timing matters. Qwen 3.6 was barely announced when this test landed, which means the community was already prodding it before the press releases had dried. Kebab prompts earn their reputation not through complexity but through the gap they expose between a model that pattern-matches "understood" and one that actually follows through — multilingual reasoning, culturally grounded tasks, the kind of instruction that looks solved until it isn't. The test itself is the story here, as much as the results. One user, one April afternoon, one thread. That's where we are.

If confirmed, here is what this means. The Qwen family — already a serious presence in the open-weights space — is being stress-tested in real time by the community that actually deploys these models, not by benchmark suites designed to flatter. If Qwen 3.6 holds up under Kebab conditions against Gemma 4, that's a signal Alibaba is shipping competitive capability fast enough to outpace formal evaluation cycles. If Gemma 4 edges ahead, it suggests Google's recent push on practical instruction-following is landing where it matters — in the hands of people who run local inference and don't forgive regressions. Either outcome puts pressure on the next tier of model releases to treat culturally grounded, multilingual evaluation as table stakes rather than a nice-to-have. The second-order effect is simpler: the community doing this work before the labs do it themselves is now a normal part of the release cycle, not an exception.

Watch for follow-up threads replicating the Kebab Test with standardised prompts, or for either Google or Alibaba to reference community benchmarks in their own documentation — that's when a single LocalLLaMA post becomes part of the official record.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →