Gemma 4 12B

Name: Gemma 4 12B
Author: Google

Google · Gemma 4 · released 2026-06-03 · data checked 2026-11-01

Gemma 4 12B is a 11.9B-parameter dense model from Google (United States), released 2026-06-03 under the Apache-2.0 license. It scores 56.9 on the Small LM Index, #6 of 26 small models. It needs about 8.9 GB of memory at 4-bit with an 8K-token context.

Small LM Index

56.9

#6 of 26 scored small models

Likely 53.2–60.6 · from 5 of 6 core benchmarks

Parameters

11.9B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

256K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best small model on that benchmark. Sources: 7 reported by Google. How we label sources.

MMLU-Pro Lab-reported Knowledge77.2

Class best: Qwen3.5 9B 82.5

GPQA Diamond Lab-reported Knowledge78.8

Class best: Qwen3.5 9B 81.7

Humanity's Last Exam Lab-reported Reasoning5.2

Class best: Falcon H1R 7B 11.1

AIME Lab-reported Math · AIME 2026, no tools77.5

Class best: Apriel 1.6 15B Thinker 88.0

LiveCodeBench Lab-reported Coding · LiveCodeBench v672.0

Class best: Granite 4.2 8B 73.2

τ²-Bench Lab-reported Agentic69.0

Class best: Qwen3.5 9B 79.1

MMMU-Pro Lab-reported Vision69.1

Class best: Qwen3.5 9B 69.2

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

SWE-bench Verified≈ 41

Specifications

Lab: Google (US · United States)
Architecture: Dense transformer
Total parameters: 11.9B
Active parameters: All (dense)
Size class: Small
Context window: 256K tokens
Input modalities: text, image, audio
Reasoning: Hybrid (thinking on/off)
License: Apache-2.0
Openness: Open weights
Memory to run: ~8.9 GB at 4-bit, 8K context (more sizes)
Hugging Face: 1.8M downloads (30 days) · 1.7K likes

Weights on Hugging Face ↗

Compare with

Gemma 4 12B vs Gemma 3 12B (12.2B, index 28.4)
Gemma 4 12B vs Ministral 3 14B (14B, index 60.0)
Gemma 4 12B vs Phi-4-reasoning-plus (14.7B, index 54.4)
Gemma 4 12B vs Phi-4 (14.7B, index 44.7)

Sources

Something wrong here? Report it.

Run it locally

Uses ggml-org/gemma-4-12B-it-GGUF (Q4_0). Swap the tag for another size: BF16, Q4_0, Q8_0.

Ollama

ollama run hf.co/ggml-org/gemma-4-12B-it-GGUF:Q4_0

llama.cpp

llama-server -hf ggml-org/gemma-4-12B-it-GGUF:Q4_0

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/gemma-4-12B-it-OptiQ-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve google/gemma-4-12B-it

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	25 GB	27 GB	42 GB
8-bit Q8_0	14 GB	16 GB	31 GB
4-bit Q4_K_M	8.6 GB	10 GB	25 GB

Check it against your GPU, Mac or phone →