Gemma 4 12B Unified

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. Q4_0 uses Google's QAT GGUF; other sizes use the published Unsloth GGUF. All totals include the multimodal projector when published separately.

Estimate formula: published weight files + 0.4027 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 9.1 GB before OS and application reserve.

Conservative FP16 estimate from the public config. Eight of 48 layers use global attention and local layers use a 1024-token sliding window.

KV method source: https://huggingface.co/google/gemma-4-12B-it/raw/main/config.json

Speed heuristic

Not benchmarked here. The calculator labels relative speed from the dense parameter count only.

Active path: 11.95B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark.

Runtimes

• Ollama
• llama.cpp
• MLX

ollama run gemma4:12b

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

• The model is multimodal and the projector adds to the download and memory footprint.
• The 256K context ceiling is not a practical default on 12-16 GB machines.

Sources

• https://huggingface.co/google/gemma-4-12B-it
• https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf
• https://huggingface.co/unsloth/gemma-4-12b-it-GGUF
• https://ollama.com/library/gemma4
• https://huggingface.co/google/gemma-4-12B-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