v2.2 · 2026-06-12
VISION / MULTIMODAL MODEL — RUNTIME REQUIREMENTS
·Official Q4_0 memory estimates cover model loading only and do not include full runtime overhead, supporting software, or KV cache.
·All variants generate text output; they do not generate image or audio output.
·Beginner summary from supplied research: Gemma 4 is Google DeepMind’s newest open Gemma family for local reasoning, coding, agents, and multimodal chat. It spans tiny edge models, a 12B audio-capable model, a fast MoE, and a stronger 31B dense model.
Gemma 4 12B Instruct
gemma-4-12b-it
Gemma 4 12B Instruct is a 11.95B multimodal model profile for local AI planning. This page records provider, license, context, quantization, hardware-fit estimates, setup hints, source links, and caveats so readers do not choose a model by name alone.
PROVIDER
Google DeepMind
FAMILY
Gemma 4
MODEL TYPE
multimodal
PARAMETERS
11.95B
MODALITIES
text, image, audio
ARCHITECTURE
Gemma 4 multimodal decoder model
CONTEXT WINDOW
256,000 tokens
TRAINING TOKENS
—
RELEASE DATE
2026-04-02
SETUP HINTS
OLLAMA
ollama run gemma4:12bLM STUDIO
Available according to the supplied June 2026 research; select a cited GGUF/MLX build and verify current app metadata before relying on hands-on setup guidance.
LLAMA.CPP
Use a cited GGUF build with current llama.cpp-compatible tooling.QUANTIZATION FILE SIZES (GGUF)
Q4_K_M
6.7 GB
Q8_0
13.4 GB
Quantization note: Q4_K_M for first local attempts unless a cited runtime recommends a different default.
RAM / VRAM ESTIMATES
MIN RAM
16 GB
COMFORTABLE RAM
32 GB
MIN VRAM
16 GB
COMFORTABLE VRAM
32 GB
These are conservative local-inference estimates, not official hardware requirements.
HARDWARE FIT
CPU only (no GPU)
no
8 GB RAM
no
16 GB RAM
limited
32 GB RAM
comfortable
8 GB VRAM
no
12 GB VRAM
no
24 GB VRAM
limited
Apple Silicon (unified memory)
limited
Hardware fit values are conservative local-inference estimates based on GGUF size plus runtime and KV-cache overhead. Actual requirements depend on context length, quantization, runtime, GPU offload, and other running apps.
BEST FOR
·Best for 32GB laptops or desktops, especially Apple Silicon and consumer GPUs.
AVOID IF
·Avoid for tiny 8GB machines needing full-quality long context, or projects requiring non-text outputs today.
CAVEATS
·Official Q4_0 memory estimates cover model loading only and do not include full runtime overhead, supporting software, or KV cache.
·Long 128K and 256K contexts can require much more RAM or VRAM than the base model-size number suggests.
·Audio input is available only on E2B, E4B, and 12B, not on 26B-A4B or 31B.
·All variants generate text output; they do not generate image or audio output.
·The 26B-A4B MoE activates only 3.8B parameters per token but still must load all parameters into memory.
·Use the official Apache 2.0 license and Google policy links for legal review before treating commercial use as cleared.
·Commercial-use status from the supplied research: permitted
·Beginner summary from supplied research: Gemma 4 is Google DeepMind’s newest open Gemma family for local reasoning, coding, agents, and multimodal chat. It spans tiny edge models, a 12B audio-capable model, a fast MoE, and a stronger 31B dense model.
SOURCE URLS
FIELD EVIDENCE
parameterSizeBai.google.dev/gemma/docs/core/model_card_4
activeParametersBhuggingface.co/google/gemma-4-E2B-it
architecturehuggingface.co/google/gemma-4-E2B-it
contextWindowTokensai.google.dev/gemma/docs/core/model_card_4
releaseDateai.google.dev/gemma/docs/core
hfGgufRepos—
q4FileSizeGbai.google.dev/gemma/docs/core/model_card_4
q5FileSizeGb—
q8FileSizeGbai.google.dev/gemma/docs/core/model_card_4
setupHintsollama.com/library/gemma4
languageshuggingface.co/google/gemma-4-E2B-it
supportsToolshuggingface.co/google/gemma-4-E2B-it
reasoningTunedhuggingface.co/google/gemma-4-E2B-it
codeTunedhuggingface.co/google/gemma-4-E2B-it
embeddingModelhuggingface.co/google/gemma-4-E2B-it
visionModelhuggingface.co/google/gemma-4-E2B-it
trainingTokens—
LANGUAGES
Unknown
CAPABILITIES
vision / image input