Gemma 4 31B

What this model is

The architecture is dense. Every parameter is part of the active path, so memory and speed scale with the full model rather than a smaller expert subset.

Memory and quantization

Decimal GB. Sizes combine the published Unsloth GGUF weight file with the multimodal projector required for image input.

Estimate formula: published weight files + 1.1744 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 22.2 GB before OS and application reserve.

Conservative FP16 estimate from the public config. Ten of 60 layers use global attention and local layers use a 1024-token sliding window.

KV method source: https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json

Speed heuristic

Not benchmarked here. Dense 30.7B inference activates the full model on every token.

Active path: 30.7B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark.

Runtimes

• Ollama
• llama.cpp
• MLX

ollama run gemma4:31b

License restrictions

Apache 2.0; the separate Gemma Prohibited Use Policy and intended-use statement still apply.

This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.

Known issues

• A Q4 build plus KV cache is a workstation-class workload, not a typical 16 GB laptop fit.
• The GGUF package needs the multimodal projector for image input.

Sources

• https://huggingface.co/google/gemma-4-31B-it
• https://huggingface.co/unsloth/gemma-4-31B-it-GGUF
• https://ollama.com/library/gemma4
• https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json

This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.

Check a specific machine