Text Gen · Model Guide · 5 min read

Gemma 4 31B

Google's most capable single-GPU Gemma — frontier-adjacent quality that fits comfortably in 24 GB.

Basic Information

Gemma 4 31B is the largest current entry in Google's open Gemma family: a dense ~31-billion-parameter model that runs on a single 24 GB consumer GPU at sensible quantizations. The Gemma line has a reputation for outsized instruction-following and writing quality per parameter, and the 4th generation continues that — strong multilingual chat, solid coding, and unusually good behavior for its class.

Being dense rather than mixture-of-experts, every parameter works on every token: memory needs scale straight from the parameter count, and speed is predictable. Google ships it in the same open-weights style as previous Gemmas — downloadable, runnable in Ollama/llama.cpp, and fine-tunable.

Best use cases

• A single-GPU daily-driver assistant: chat, summarization, writing help
• Coding support on workstation hardware without cluster budgets
• Privacy-first deployments — it runs fully offline after download
• Fine-tuning experiments at a scale hobbyist hardware can actually handle

Computational Requirements

The sweet spot for one modern 24 GB card: Q4 puts the weights around 14–15 GB, leaving real headroom for context on the same card.

• On a 16 GB card use Q3/Q2 variants; on 24 GB, Q4 with an 8–16K context is comfortable.
• Check the exact fit for your card and context length in the compatibility panel above.

Commercial license

Gemma Terms of Use: commercial use is permitted, with conditions — a prohibited-use policy, restrictions on redistributing the weights outside the license terms, and Google's standard indemnity carve-outs. Small-team friendly, but it is a bespoke license, not Apache/MIT — read it before shipping a product around it.