by Google DeepMind

Gemma 4 12B review, pricing and limits

The mid-sized, laptop-runnable member of the Gemma 4 family, built to read text, images, audio and video on a single 8GB GPU.

Best for
On-device multimodal agents
Not for
Ceiling-level agentic coding
  • Available
  • Open source

Should you use Gemma 4 12B?

Developers building on-device or self-hosted multimodal agents get a model that fits about 6.6GB of VRAM and skips per-token API bills entirely. It replaces separate vision-plus-audio encoder pipelines with one encoder-free model, though teams chasing top-tier agentic coding should size up to Gemma 4 31B instead.

Best for

  • On-device multimodal agents Encoder-free design keeps latency low for image, audio and video input on consumer hardware.
  • Self-hosted document and screen understanding Native OCR, chart reading and UI understanding without a separate vision encoder pipeline.
  • Cost-sensitive agentic tool-use apps Free self-hosting under Apache 2.0, or a fraction of a cent per request via SiliconFlow's hosted API.

Not for

  • Ceiling-level agentic coding Its own 31B sibling scores 89.2% on AIME 2026 versus unverified numbers for the 12B; go bigger if raw reasoning is the priority.
  • Voice-to-voice assistants Audio input only, no native audio output, so a separate TTS model is still required.

Gemma 4 12B pricing

Pricing model
free (self-hosted, open weights)
Free tier
Fully free to download and self-host under Apache 2.0; no usage caps since it runs on your own hardware.

Cost examples

Self-host on a single RTX 4090 or M2/M3 MacBook Pro
$0.00
1M input / 200K output tokens via SiliconFlow hosted API
$0.16
TierRate
Self-hosted (open weights)Free under Apache 2.0. Requires your own GPU (about 6.6GB VRAM at Q4_K_M) or CPU (slower).
SiliconFlow hosted APIThird-party managed inference, not sold directly by Google.

Gemma 4 12B is free to self-host under an open license: there is no per-token fee and no rate limit when you run it on your own GPU. Google does not sell hosted inference for Gemma directly. Third-party host SiliconFlow prices its managed API for the model at $0.10 for every 1M input tokens and $0.30 for every 1M output tokens, roughly $0.12 blended, as of July 2026.

Self-hosting requirements

Min vram gb
6.6
Self-hosted
Yes
Recommended GPU
8GB+ consumer GPU (RTX 3060/4060 class) or 16GB unified-memory Apple Silicon

What Gemma 4 12B can do

Gemma 4 12B scores 78.8% on GPQA Diamond, Google DeepMind's mid-sized open-weight multimodal model released June 3, 2026. Its encoder-free architecture reads text, images, audio and video directly, running on a single consumer GPU under an Apache 2.0 license with no per-token fee.

What goes in and what comes out

Reads

  • text
  • image
  • audio
  • video

Writes

  • text
  • tool-calls

Capabilities

  • Audio inputSupported
  • Function callingSupported
  • Structured outputSupported
  • Tool useSupported
  • Video inputSupported
  • VisionSupported
  • Audio outputNot supported
  • Code executionNot supported
  • Web browsingNot supported
Reasoning modes
standard

Key features

  • Encoder-free multimodal input

    A 35M-parameter linear embedder projects raw image patches and audio directly into the model, skipping the heavyweight vision/audio encoders most multimodal models rely on.

  • 256K token context window

    262,144 tokens shared across the whole Gemma 4 family, enough for full codebases or long video transcripts on a 12B-class model.

  • Runs on 8GB consumer GPUs

    About 6.6GB VRAM at Q4_K_M quantization, with official Google QAT checkpoints that hold quality close to bf16.

  • Native function calling

    Structured tool-use support built in for agentic workflows, without a separate fine-tune or adapter.

  • Video and audio understanding

    Processes video as sampled frame sequences and accepts native audio input, alongside text and image, in one unified model.

Known limitations

  • Google has not published an official SWE-bench or MMLU-Pro score for the 12B specifically; the reported GPQA Diamond (78.8%) and DocVQA (94.9%) figures are third-party, not vendor-confirmed.
  • No native audio output, only audio input, so it cannot be used alone for voice-to-voice assistants.

Who it is for

  • Developers building on-device or self-hosted multimodal agents
  • Teams that want to avoid per-token API costs by running inference on their own GPUs
  • Researchers fine-tuning a permissively-licensed multimodal base model
  • Startups needing document, screen and chart understanding without a Vertex AI or OpenAI bill

Supported languages

  • English
  • Spanish
  • French
  • German
  • Japanese
  • Chinese
  • Arabic
  • Hindi
  • Portuguese
  • Korean

Gemma 4 12B benchmarks

Gemma 4 12B against its nearest rivals

GPQA Diamond % correct, higher is better

Standout result

262,144

token context window, the shared spec across the entire Gemma 4 family

Source: Google AI for Developers Gemma 4 model overview, June 2026

Where it sits

Mid-pack on GPQA Diamond (rank 42 of 63), and one of 90 whose vendor states it does not train on customer data.

MeasureThis modelRankMedian
GPQA Diamond% correct78.8%#42 / 63peer median 87.8%

Ranks are against GA models on Hokai that publish the same figure; ties share a rank.

Benchmark scores

  • DocVQA vendor-reported · 03 Jun 202694.9
  • GPQA Diamond vendor-reported · 03 Jun 202678.8%PhD-level science questions that are hard to search for, % correct.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Strengths and limits

Strengths

  • Carries no per-token license fee and drops the Google-specific usage restrictions that came with the older Gemma Terms of Use.
  • Encoder-free architecture that Google says cuts multimodal inference latency compared with encoder-heavy rivals of similar size.
  • Fits on an 8GB GPU or a 16GB unified-memory laptop thanks to official quantization-aware training checkpoints.
  • Runs on an 8GB consumer GPU at Q4_K_M quantization (about 6.6GB VRAM), no data center needed.
Show 2 more
  • Encoder-free design projects images and audio directly into the model, cutting multimodal latency versus encoder-heavy rivals of similar size.
  • 256K token context window shared across the whole Gemma 4 family, letting a 12B model handle full codebases or long video transcripts.

Limits

  • No official SWE-bench, AIME or MMLU-Pro score published for the 12B; only third-party GPQA Diamond and DocVQA figures exist so far.
  • Audio input only, no audio output, so voice assistants still need a separate TTS model.
  • Meaningfully behind its own 31B sibling and closed frontier models like Gemini 3.1 Pro on ceiling reasoning tasks.

Gemma 4 12B alternatives

ModelLabContextIn / out per 1MReads and writes
Gemma 4 12B This pageGoogle262K--text, image, audio, video → text
Gemma 4 31BPick Gemma 4 31B if you need its higher reasoning benchmarks and can run it on data-center hardware; pick Gemma 4 12B to fit an 8GB GPU.Google262K--text, image → text
Gemini 3.1 ProPick Gemini 3.1 Pro for a higher benchmark ceiling and native audio output; pick Gemma 4 12B for free self-hosting on your own hardware.Google1M$2 / $12text, image, audio, video, pdf → text
Hy4 flashTencent's Hy4 flash, released 16 November 2026, is a lower-cost multimodal model that reads text and images across a 1M-token window.Tencent1M$0.07 / $0.281text, image → text

Gemma 4 12B specs

Context window
262K tokens
Tokens per second
45 tok/s
Mixture of experts
No
Parameters
12B
Architecture
Transformer (dense, decoder-only, encoder-free multimodal)
VRAM requirements
~6.6GB at Q4_K_M quantization; fits an 8GB consumer GPU or 16GB unified-memory laptop

Training data

Web documents, code, images and audio spanning 140+ languages, filtered for personal information, CSAM and quality/safety issues per Google's data-handling policy. Training data cutoff January 2025.

Weight formats

  • safetensors (bf16)
  • GGUF (Q4_0 QAT)
  • Q4_K_M and other llama.cpp quantizations

API endpoints

NameEndpointAuth
Google Vertex AI Model Gardenhttps://cloud.google.com/vertex-aiGCP IAM
AWS Bedrockhttps://aws.amazon.com/bedrock/AWS IAM
SiliconFlowhttps://artificialanalysis.ai/models/gemma-4-12b/providersAPI Key
Hugging Face Inferencehttps://huggingface.co/google/gemma-4-12B-itAPI Key / self-hosted

SDKs

  • Python
  • TypeScript
  • Go
  • Java

Regions

  • Any region you self-host in
  • us (Vertex AI Model Garden)
  • global (AWS Bedrock)

Prompting patterns

Show 2 prompting patterns
  • Interleave text and images directly

    Mix <image> tokens with instructions in a single prompt rather than a separate vision call.

    The encoder-free design was built for native interleaved multimodal input.

  • Use the official QAT GGUF checkpoint for local serving

    Load google/gemma-4-12B-it-qat-q4_0-gguf instead of a community re-quant.

    Google's quantization-aware training preserves bf16-level quality that generic post-hoc quantization loses.

Known quirks

Show 3 known quirks
  • No official 12B reasoning benchmarks yet

    Google's launch materials published SWE-bench, AIME and MMLU-Pro numbers for the 31B model but not the 12B, leaving GPQA Diamond and DocVQA as the only widely-cited 12B scores, and those come from third-party write-ups.

    Workaround: Run your own eval against your target task before committing, and treat published 78.8% GPQA Diamond / 94.9% DocVQA figures as directional, not vendor-guaranteed.

  • Community GGUF re-quants can break tool calling

    Early community-uploaded GGUF conversions shipped with an outdated chat template that mishandled function-call formatting until Google's official template fix landed in June 2026.

    Workaround: Pull the official google/gemma-4-12B-it-qat-q4_0-gguf checkpoint or a repack dated after the June 2026 template fix, not an older community mirror.

  • Audio support is size-gated within the family

    Only the E2B, E4B and 12B checkpoints accept audio input; the larger 26B-A4B and 31B Dense siblings do not, which surprises teams assuming bigger means more modalities.

    Workaround: Stay on the 12B (or smaller) checkpoint if audio input is required, and pair it with the 31B via a routing layer if you also need its reasoning ceiling.

Gemma 4 12B facts

Trains on your data
No
Licence
Apache License 2.0 · Open source
Safety posture
Balanced
Released
03 Jun 2026
Current version
gemma-4-12b-20260603
Knowledge cutoff
01 Jan 2025
Family
Gemma 4
Variant
Multimodal
Deployment
Self-hosted · Google Vertex AI · AWS Bedrock · Ollama · LM Studio · vLLM
Data retention
Self-hosted or third-party-hosted; no data retention by Google since Gemma is not served through a Google-operated inference API.

Content restrictions

  • Filtered during pretraining for CSAM, sensitive personal data, and low-quality content per Google's Gemma data-handling policy.
  • Apache 2.0 license carries no Google-specific usage restrictions (unlike the earlier Gemma Terms of Use), so content moderation is left to the deployer.

Safety controls

Rlhf
Yes
Red teaming
documented in Google's Gemma 4 model card
System prompt
Yes
Constitutional AI
No
Content filter API
No
Moderation endpoint
No

Data governance

Trains on inputs
No
Data residency options
fully self-hosted, any region
Eu AI act classification
Open-weight general-purpose AI model; obligations fall on the deploying organization, not Google, for self-hosted use.

Version history

  • gemma-4-12b-20260603 · 03 Jun 2026

    12B added as the mid-sized member of the Gemma 4 family, months after the April 2 launch of E2B/E4B/26B-A4B/31B.

  • gemma-4-family-20260402 · 02 Apr 2026

    Initial Gemma 4 family launch (E2B, E4B, 26B-A4B MoE, 31B Dense) under Apache 2.0, replacing the older Gemma Terms of Use.

What changed

  • Release Gemma 4 12B added to the Gemma 4 family (gemma-4-12b-20260603) source
  • Release Gemma 4 family launches under Apache 2.0 (gemma-4-family-20260402) source

Questions about Gemma 4 12B

How much does Gemma 4 12B cost in 2026?

Nothing, if you self-host Gemma 4 12B on your own GPU: it's open-weight with no license fee. SiliconFlow's managed endpoint bills input tokens at $0.10 per million and output tokens at $0.30 per million, about $0.12 blended, which works out to roughly $0.16 for a 1M-input, 200K-output run. Google itself does not sell Gemma inference directly.

What do you get on Gemma 4 12B's free tier?

The whole model, with no tier at all: self-hosting Gemma 4 12B carries no license fee, no seat limits, and no expiring trial since it runs on your own hardware. It needs roughly enough VRAM to hold a heavily quantized 12B-parameter model, an 8GB consumer GPU or a similarly modest Apple Silicon Mac is generally enough. The only ongoing cost is your own electricity, not a Google or third-party bill.

What are the best alternatives to Gemma 4 12B?

The closest competitors are Gemma 3 12B, Llama 4 Scout, and Qwen3 14B, similarly sized open-weight models with multimodal or agentic strengths of their own. Choose Llama 4 Scout if you specifically need Meta's tool ecosystem, or Qwen3 14B if multilingual coverage matters more than encoder-free latency. Gemma 4 12B's edge is combining text, image, audio, and video input in one checkpoint without a separate vision or audio encoder.

Is Gemma 4 12B better than Gemma 4 31B?

Not on reasoning: Gemma 4 31B is the family's reasoning-focused sibling, and Google's own benchmarks put it at 89.2% on AIME 2026, a tier the 12B has no vendor-confirmed equivalent for yet. Both share the same context window and encoder-free multimodal input, but the 31B needs data-center-class hardware while the 12B fits an 8GB consumer GPU. Pick the 31B for agentic coding and math-heavy work; pick the 12B for a laptop-runnable multimodal agent.

What does it take to start using Gemma 4 12B?

Pull the official google/gemma-4-12B-it-qat-q4_0-gguf checkpoint from Hugging Face rather than an older community re-quant, since early community conversions shipped a chat template that broke function-calling formatting. Load it in Ollama, LM Studio, or llama.cpp; on an 8GB GPU it runs at interactive speed with no extra configuration. For managed hosting instead of local setup, Google Vertex AI Model Garden and AWS Bedrock both list Gemma 4 12B, and SiliconFlow offers pay-per-token API access without any GPU at all.

About Gemma 4 12B

Gemma 4 12B is an open-weight multimodal language model built by Google DeepMind, released June 3, 2026 as the mid-sized addition to the Gemma 4 family that launched April 2, 2026 alongside the E2B, E4B, 26B-A4B mixture-of-experts, and 31B dense variants. It sits between the tiny edge-focused E4B and the larger 26B/31B models, aimed at developers who need a single laptop-runnable model that reads text, images, audio, and video without shipping four separate pipelines. It uses a dense, decoder-only Transformer built on the same backbone as Gemma 4 31B Dense, with roughly 12 billion parameters.

Read the full background

The defining architectural choice is going encoder-free. Earlier multimodal stacks bolt on a heavyweight vision encoder (often 500M+ parameters) and a separate audio encoder, then fuse their embeddings into the language model. Gemma 4 12B drops both: a 35M-parameter linear vision embedder projects raw 48x48 image patches directly into the model's hidden dimension in a single matrix multiplication, with spatial position added through factorized X/Y coordinate lookup tables instead of a full attention stack. Audio gets the same lightweight-projection treatment, which is why Google reports lower multimodal inference latency than the encoder-heavy designs used in Gemma 3 and most competing open models.

On benchmarks, independent write-ups (Techsy, BuildFastWithAI, AICybr) report Gemma 4 12B scoring roughly 78.8% on GPQA Diamond and 94.9% on DocVQA. Google has not published an official SWE-bench or MMLU-Pro score for the 12B variant specifically; those numbers exist for the larger 31B sibling, which posts 89.2% on AIME 2026 and 80.0% on LiveCodeBench v6, so treat the 12B reasoning figures as third-party reported rather than vendor-confirmed until Google ships a dedicated model card update.

Context window matches the rest of the Gemma 4 family, which lets a 12B-class model handle full codebases, long PDFs, or hour-plus video transcripts without chunking. Google has not published a separate max-output-token ceiling distinct from that shared context budget.

Modality support is the headline feature: text, image, video (processed as sampled frame sequences), and audio input, with the E2B, E4B, and 12B checkpoints all sharing audio support (the larger 26B and 31B variants do not, per Google's own comparison). Capabilities include object detection, document and PDF parsing, screen and UI understanding, chart reading, multilingual OCR, and handwriting recognition. Interleaved multimodal prompts, mixing text and images in any order, are supported natively rather than through a special API mode.

Deployment options span Hugging Face, Kaggle, Ollama, LM Studio, vLLM, SGLang, MLX, and llama.cpp for local or self-managed serving, plus Google Cloud's Vertex AI Model Garden, Cloud Run GPU, and GKE for managed cloud deployment; AWS added Gemma 4 support to Bedrock in mid-June 2026. It ships free to self-host under an open license, with third-party hosted-API rates covered under pricing. Google also ships an official multi-token-prediction (MTP) drafter model for Gemma 4 that adds speculative decoding, claiming up to 3x faster generation with no quality loss when paired with the 12B base model.

The model shipped under a plain Apache 2.0 license, a meaningful change from the earlier custom "Gemma Terms of Use" that governed Gemma 1 through 3. That older license was broadly permissive but carried Google-specific use restrictions that made some enterprise legal teams hesitate; Apache 2.0 is the same OSI-approved license used by Qwen and Mistral, meaning commercial fine-tuning, redistribution, and paid hosting are unambiguously allowed with just attribution and license-text inclusion, no derivative-weight sharing required.

Pre-training data has a cutoff of January 2025 and spans web documents, code, images, and audio across more than 140 languages, filtered for personal information, CSAM, and general safety and quality issues per Google's stated data-handling policy. Best fit is developers building on-device or self-hosted multimodal agents, such as browser automation, document pipelines, or voice-plus-vision assistants, who want to keep data on their own infrastructure.

Sources and guides

More AI Models on Hokai

  • Nemotron-Labs-SigmaNVIDIA's Nemotron-Labs-Sigma, released in November 2026 in 8B and 3B sizes, writes text by denoising whole blocks of tokens in parallel, not one at a time.
  • Qwen4 MaxAlibaba's Qwen4 Max, released 16 November 2026, is a sparse mixture-of-experts model with a 1M-token context and text, image and video input.
  • Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio.
  • Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.
  • Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.
  • Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.