# Gemma 4 vs Qwen 3.6: which one returns an empty answer under a tight token budget?

With the same everyday requests and the same 1024-token budget, which model spends the budget thinking and returns nothing?

Under the same tight budget, the model that reasons longer runs out of room before saying anything far more often. Both signals arrive as HTTP 200 — a client reading only the status code sees success either way. The cure on the Qwen side is turning reasoning off, which eliminates the failure mode on this corpus entirely.

**What was held fixed.** Same unit, backend, corpus and 1024-token budget. The operating mode is each model’s shipped reasoning behavior: gemma-4 in its default mode, Qwen3.6 with reasoning enabled — this is a comparison of defaults under pressure, not of identical settings.

| Metric | gemma-4-26B-A4B (default mode) | Qwen3.6-35B-A3B (reasoning on) |
| --- | --- | --- |
| Empty answers under HTTP 200 (answerless), share of requests | 8.3 % of requests ([lab_repeated](https://agmind.ai/claims/strix.gemma4.interactive2.c1.answerless-default/)) | 58.3 % of requests ([lab_unit_replicated](https://agmind.ai/claims/strix.qwen36.interactive2.c1.answerless-1k/)) |
| Time to first token of anything, median | 282 ms ([lab_repeated](https://agmind.ai/claims/strix.gemma4.interactive2.c1.ttft-any-token/)) | 212 ms ([lab_unit_replicated](https://agmind.ai/claims/strix.qwen36.interactive2.c1.ttft-any-token/)) |
| Time to first ANSWER token, median | 10768 ms ([lab_repeated](https://agmind.ai/claims/strix.gemma4.interactive2.c1.ttfa-default/)) | 20191 ms ([lab_unit_replicated](https://agmind.ai/claims/strix.qwen36.interactive2.c1.ttfa-thinking/)) |

## Limits

One budget, one corpus, defaults-vs-defaults: gemma-4 with reasoning suppressed was not tested, and a bigger budget changes both numbers. This measures how the shipped configurations fail, not which model is smarter.

## Related evidence

- https://agmind.ai/reports/answerless-http-200/
- https://agmind.ai/essays/dashboard-lies-both-directions/

Every number above is re-derived from sealed run bundles in CI: https://agmind.ai/claims.json (CC BY 4.0).
Source page: https://agmind.ai/compare/gemma4-vs-qwen36-answerless/