Gemma 4 31B narrows Qwen3.5 gap, cuts tokens

Qwen still looks like the better all-around benchmark winner, but Gemma 4 31B’s lower token use is the more interesting signal for developers paying real inference bills.

• –A model that is slightly behind on average but far cheaper to run can be the better choice for agent loops, long chats, and self-hosted deployments.
• –The 31B wins suggest Gemma 4 is not just a smaller, cheaper cousin; it can still take specific reasoning and knowledge tests.
• –If the token-efficiency gap holds across workloads, it can outweigh raw score deltas once you factor in latency, GPU memory, and throughput.
• –For teams choosing between Gemma and Qwen, the real decision is increasingly quality-per-token, not just leaderboard rank.