Inference model · GEMMA

Gemma 4 31B

Gemma's dense flagship in our catalog. Different inductive biases than Qwen — keep both in your evals.

Request Gemma 4 31B access Back to catalog

Gemma 4 31B brings a different family of training data and reward signals than the Qwen line. Worth keeping in your evaluation set: some prompts that Qwen handles unevenly land cleanly on Gemma, and vice versa.

When to pick it

  • Long-form writing, prose summarization, content workflows
  • Multi-lingual content where Qwen’s bias toward Chinese-heavy training shows
  • Eval coverage across families (don’t single-source your agent stack)

When to look elsewhere

  • 64K context windows → not the right model

  • Throughput-sensitive batch jobs → Gemma 4 26B A4B (MoE) is much cheaper per token