Alpha
Contact
Implement · Gemma 4

Can you run it?

Ownership levelSubstantialnone·limited·partial·substantial·fullAnalytical input C ยท 68.8/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

The weights are ungated, so no Hugging Face login or approval is needed (ev-hf-inventory). Pin an exact revision and verify the LFS checksums after download.

# Hugging Face transformers (from the model card)
from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "google/gemma-4-31B-it"      # or -26B-A4B-it, -12B-it, -E4B-it, -E2B-it
processor = AutoProcessor.from_pretrained(MODEL_ID, revision="<pinned-commit-sha>")
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, revision="<pinned-commit-sha>", dtype="auto", device_map="auto"
)
# Ollama (tags e2b, e4b, 12b, 26b, 31b)
ollama run gemma4

For llama.cpp or LM Studio use the official -qat-q4_0-gguf repositories; for vLLM or SGLang use -qat-w4a16-ct (ev-docs-overview, ev-ollama). The model card says to use the latest Transformers release (ev-card-31b).

Hardware & VRAM requirements

Google's overview page gives approximate weight-load memory including 20 percent overhead (ev-docs-overview):

SizeBF168-bitQ4_0
31B dense69.9 GB34.9 GB17.5 GB
26B A4B (all 25.2B parameters resident)57.7 GB28.8 GB14.4 GB
12B Unified26.7 GB13.4 GB6.7 GB
E4B17.9 GB8.9 GB4.5 GB
E2B11.4 GB5.7 GB2.9 GB

The E2B and E4B also have mobile builds (E2B 1.1 GB, E4B 2.5 GB in the table's mobile column). These figures cover weights only. The KV cache at 128K to 256K context comes on top, and fine-tuning needs far more. The launch blog says the 31B BF16 weights fit one 80 GB H100 (ev-blog). The 26B A4B runs at roughly 4B-parameter speed but needs memory for all 26B.

Serving stacks

StackUsed for
TransformersReference loading, prototyping, fine-tuning
vLLM / SGLangHigh-throughput serving with the official w4a16 compressed-tensors builds
llama.cpp / LM StudioLocal and CPU use with the official q4_0 GGUF builds
OllamaSingle-command local runs for all five sizes (ev-ollama)
LiteRT-LMMobile builds named by Google (ev-docs-overview)

Speculative decoding uses the matching -assistant drafter with a quantised main model (ev-docs-overview). The launch blog lists further day-one stacks (MLX, NVIDIA NIM, Unsloth, Docker and others). That list is a publisher claim and is unchecked (ev-blog).

Safe-deployment controls & Deployment Ceiling

The model-specific gate rows follow the safe-deployment playbook:

  • Pull from the verified google organisation only, pin the revision and verify checksums. Prefer Google's own QAT repositories to community quants.
  • Deploy an instruction-tuned checkpoint. Use the card's sampling settings (temperature 1.0, top_p 0.95, top_k 64) and strip thinking content from history.
  • Add your own guardrails and a guard classifier. No Gemma 4 guard model has been identified, and the publisher safety evidence is qualitative (ev-tech-report).
  • No Prohibited Use Policy binds the weights, so enforce your own at the application layer.
  • Test image, audio and video paths separately. Limits are 30 seconds of audio and 60 seconds of video at one frame per second (ev-card-31b).
  • Record in the AI-BOM that the training summary states user data from Google products and Gemini 3 synthetic data were used (ev-eu-summary).

Deployment Ceiling: T2 (customer-facing, human-reviewed), conditional. With an instruction-tuned checkpoint, your own guardrails and per-language, per-modality evaluation, Gemma 4 reaches customer-facing, human-reviewed use. T3 and above need a deterministic action-policy engine and independent safety evidence that does not yet exist for this release.

Available quantizations

Google publishes official quantisation-aware-trained checkpoints in the google organisation, all Apache-2.0 and ungated (ev-docs-overview, ev-hf-inventory, ev-tech-report):

  • -qat-q4_0-gguf for llama.cpp and LM Studio (all five sizes);
  • -qat-w4a16-ct compressed-tensors for vLLM and SGLang (E2B, E4B, 12B, 31B);
  • -qat-mobile-ct and -qat-mobile-transformers for edge devices;
  • -qat-q4_0-unquantized half-precision QAT weights for converting to other formats such as MLX, plus -assistant drafters for speculative decoding.

Test quality for your language and task after any quantisation.

Fine-tuning & adaptation

Fine-tune the base or -it checkpoints with Transformers or Google's JAX library (ev-gemma-lib). The 12B Unified model is encoder-free, so the whole model can be tuned in one pass (ev-card-31b). Fine-tuning needs the right end token: base models end with <eos>, instruction-tuned models with <turn|> (ev-tech-report).

Modifications may be licensed on the deployer's own terms, but redistribution requires shipping the Apache-2.0 text, marking changed files and keeping notices (ev-license). A Gemma 3 derivative, by contrast, has to carry the Gemma Terms and the Prohibited Use Policy (see Assess, license item). Placing a derivative on the EU market may make the deployer its provider under the AI Act.

API / OpenAI-compatible integration

The named servers (vLLM, SGLang, Ollama, LM Studio, llama.cpp server) expose chat endpoints, but OpenAI-compatible serving of Gemma 4 is not documented in a primary document, so coverage is partial. Test tool-call round trips and thinking-mode output through the gateway. The chat template drives tool calls and thinking blocks, and it has had several fixes since release (ev-hf-scan). Keep thinking content out of conversation history except on tool-call turns (ev-card-31b).

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Moderate

Operational readiness is good (score 4), but the safety evidence is publisher-only and qualitative (score 3) and no guard model has been identified, so the deployer supplies the safety layer (ev-tech-report, ev-docs-overview).

How this scores (AOI sub-dimensions)
Operational4/5how practical it is to run, serve and maintain in productionStandard safetensors on Transformers, with Google documenting official quantised routes for llama.cpp, vLLM and SGLang, drafter checkpoints for speculative decoding, a memory table per precision and 128K to 256K context.
Safety3/5whether misuse risks are evaluated and guardrails are providedInstruction-tuned checkpoints with a publisher safety evaluation and stated policies (child safety, dangerous content, sexually explicit content, hate speech, harassment), documented risks and mitigations, and a Responsible Generative AI Toolkit pointer.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted weights under Apache-2.0 with no clawback of rights, so prompts and outputs stay on the deployer's infrastructure. The licence carries no telemetry or remote-restriction clause (ev-license).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionUse of these weights is owned outright: plain Apache-2.0, no gate, no use policy and no remote shut-off, which the older Gemma Terms of Use did have. What the weights are is not owned: the training data is closed, Google states that user data from its products went into training, and safety evidence is publisher-only. Adopt the instruction-tuned checkpoints behind your own guardrails and check the licence per checkpoint, because Gemma 1 to 3n and the Gemma 3 derivatives you may already run remain under the old terms.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Licenceread2026-11-19
The Gemma 4 licence page (linked as license_link by every Gemma 4 model card) is the Apache License 2.0.
Documentationread2026-11-19
The Hugging Face API (anonymous) shows the google organisation as verified.
Model cardread2026-11-19
The Gemma 4 model card (README of google/gemma-4-31B-it, same text in the other cards, plus the gemma.ai model card page) lists five sizes: E2B, E4B, 12B Unified, 26B A4B and 31B.
Documentationread2026-11-19
The Gemma 4 overview page gives a memory table per precision for each size and routes official QAT checkpoints: q4_0 GGUF for llama.cpp and LM Studio, w4a16 compressed-tensors for vLLM and SGLang, mobile builds, unquantised QAT weights and matching -assistant drafters.
Technical_reportread2026-11-19
The Gemma 4 Technical Report (arXiv 2607.02770 v2, 24 Jul 2026) states that the models are released under Apache 2.0.
Documentationread2026-11-19
Google's EU AI Act Article 53(1)(d) Public Summary of Training Content for the Gemma 4 family is v1, last updated 31 July 2026, from provider Google Ireland Limited.
Terms of serviceread2026-11-19
The Gemma Terms of Use (last modified April 1, 2026) apply only to the models in their Appendix (Gemma 1, 1.1, 2, 3, 3n and siblings; Gemma 4 is not listed) and state 'For Gemma 4 terms, see the Gemma 4 license'.
Terms of serviceread2026-11-19
The Gemma Prohibited Use Policy (last modified February 21, 2024) lists restricted uses for 'Gemma or Model Derivatives' and is bound into a contract only through section 3.2 of the Gemma Terms of Use.
Vendor announcementread2026-11-19
The launch blog (2 April 2026) says Gemma 4 is 'released under a commercially permissive Apache 2.0 license'.
Third-party analysisread2026-11-19
Hugging Face platform data for google/gemma-4-31B-it shows scansDone true with no files with issues, safetensors BF16 parameters 31,273,088,876, an LFS sha256 for each weight shard and no signature files.
Third-party analysisread2026-11-19
The Decoder (2 April 2026) reports Gemma 4 shipping under Apache 2.0 'for the first time' and says earlier Gemma versions used a more restrictive Google proprietary licence.
Third-party analysisread2026-11-19
The Artificial Analysis page for Gemma 4 31B (Reasoning) gives an Intelligence Index v5.0.1 score of 4 and 'Intelligence #14 / 15', and describes the model as at the lower end among comparable open-weights models (median 10).
Third-party analysisread2026-11-19
The European Commission page on the General-Purpose AI Code of Practice (last update 7 October 2026) lists Google among the signatories; it does not say which chapters.
Third-party analysisread2026-11-19
Ollama library page for gemma4: tags e2b, e4b, 12b, 26b and 31b; capability tags vision, tools, thinking and audio; 29.5M downloads; run command 'ollama run gemma4'.
Documentationread2026-11-19
google-deepmind/gemma README and LICENSE: a JAX library to use and fine-tune Gemma with Gemma 4 checkpoint paths; code under Apache-2.0; the library is not the pre-training pipeline; issues go to the repository's GitHub tracker.
Securityread2026-11-19
Google's organisation-wide GitHub SECURITY.md says to report vulnerabilities at g.co/vulnz and that the Google Security Team processes a report within a day and responds within a week.
Third-party analysisunverified2026-11-19
The ADAPT Centre AI Accountability Lab evaluation of the Gemma 4 public training summary is referenced by a secondary blog.