Gemma family

Open-weight family of general-purpose Google models, including Gemma 4 multimodal and text variants.

Score calculation

(4 + 7 + 8 + 4 + 3) ÷ 5 = 5.2

5 of six dimensions rated. Unrated dimensions and BDS status are excluded.

Setup details

This entry covers the general-purpose Gemma family, especially the current Gemma 4 core releases. Organization-wide privacy, labour, energy, military, and Israel-Palestine relationships are not inferred from this model entry. Specialized Gemma releases are noted separately from the core general-purpose family.

The official releases catalog lists Gemma 4 (E2B, E4B, 12B, 26B A4B, and 31B) and a June 3, 2026 Gemma 4 12B Unified release. The catalog also lists EmbeddingGemma 2 on October 6, 2026; that is a specialized embedding release, not a newer general-purpose core checkpoint. Checked 2026-10-23.

Energy

↑

the Gemma 4 card describes small-model and on-device efficiency, but gives no training electricity, GPU-hours, emissions, water, or standardized per-inference energy figures. Hardware footprint claims are not measured energy.

Reasoning & sources

The Gemma 4 card’s on-device efficiency claims describe deployment capability rather than measured training or request energy. Without energy, emissions, and water boundaries, the impact cannot be compared; measure the specific local or hosted workload.

No measured training or inference energy, emissions, or water disclosure was found in the checked Gemma 4 card or technical report.

Training rights

↑

The model card describes a broad mixture of web documents, code, mathematics, images, and audio, with a January 2025 data cutoff and filtering steps. It does not provide a source-level training-data manifest or a legal rights audit.

Reasoning & sources

The Gemma 4 card identifies source categories, modalities, filtering, and a January 2025 cutoff, but no source-level manifest or rights audit. Users can understand the broad dataset scope, but cannot verify individual permissions or creator opt-out coverage.

Some dataset and filtering description is published, but source attribution and rights clearance are not documented.

Corpus composition by source, consent and licensing status, creator opt-outs, and independent rights review remain undisclosed.

Privacy

↑

Gemma 4 weights can be downloaded and run locally or on-device, allowing an operator to keep prompts on infrastructure they control. This is a deployment option, not a model-level privacy guarantee; hosted Google services have their own data handling terms.

Reasoning & sources

Downloadable Gemma weights can run on-device or on infrastructure controlled by the operator, reducing exposure to a hosted service when configured locally. This does not remove local logging or application risks; hosted Google use requires separate service-policy review.

Local deployment can reduce data disclosure to a service provider, with safeguards dependent on operator choices.

Privacy depends on the chosen runtime, host, telemetry, and application; the model card does not audit deployments.

Openness

↑

Gemma 4 provides downloadable pre-trained and instruction-tuned weights under Apache 2.0, with model cards and a technical report. The training corpus itself is not released as a source-level dataset.

Reasoning & sources

Apache 2.0 weights, model cards, and a technical report give broad access to deployment and technical inspection. Because the training corpus is not released, the family is not fully reproducible despite its permissive model license.

Permissive license, weights, and technical documentation are available; source training data remain closed.

Training data, full data processing pipeline, and a fully reproducible training run are not publicly supplied.

Worker treatment

↑

Google’s supplier code applies labor and human-rights requirements across direct, temporary, migrant, contract, and subcontracted workers; it allows audits, requires remediation, and sets a 60-hour weekly limit with stated exceptions. This is a supplier policy, not an audit of AI data workers.

Reasoning & sources

Google publishes broad supplier protections and audit rights, but this review found no worker-level results for AI data work. This is relevant when deciding whose business to support, while pay and working conditions for the team behind Gemma family remain unestablished. Ask for current worker outcomes for that team and its suppliers.

The Gemma card reports human safety evaluations but not who performed them, how many workers were involved, or their terms. Safety evaluation evidence does not establish fair labor conditions; request contractor and evaluator disclosures.

Google publishes broad supplier protections and audit rights, but this review found no worker-level results for AI data work. This rating describes the company relationship or supplier policy, not a finding about this particular product’s use.

Working conditions for the people building or supporting Gemma family are not established by this company-wide record. No current AI-annotation-specific outcomes, public audit findings, or remediation data for Vertex-related data work were established.

Government & military

↑

Google documents access for Department of War personnel and classified defense workflows. Separately, Israel’s procurement administration identifies Google as a Nimbus public-cloud provider; AFSC and Who Profits describe military/security use, while the Guardian reports contract terms limiting service withdrawal. These sources do not establish Google’s role in a particular attack.

Reasoning & sources

Google documents U.S. Department of War and classified defense workflows; Israel’s procurement record identifies Google as a Nimbus cloud provider, and independent research/reporting describes military or security use and limits on withdrawal. The reviewed evidence does not attribute a specific harmful operation to Google. This is relevant when deciding whose business to support. Deployment of Gemma family or this particular model release under those contracts remains unestablished; consider the company relationship separately from the product’s actual use.

The Gemma card and prohibited-use policy do not identify a Gemma-specific military contract or deployment, and a use policy cannot prove what customers actually do. No model-specific relationship is established; assess any separate Google relationship only at its documented scope.

Google documents U.S. Department of War and classified defense workflows; Israel’s procurement record identifies Google as a Nimbus cloud provider, and independent research/reporting describes military or security use and limits on withdrawal. The reviewed evidence does not attribute a specific harmful operation to Google. This rating describes the company relationship or supplier policy, not a finding about this particular product’s use.

Deployment of Gemma family under these contracts is not established. The source does not disclose contract terms or evaluate downstream deployment outcomes.

BDS status

↑

Google / Alphabet is named as a pressure target in the BNC’s official campaign sources. The product is developed or supplied by Google / Alphabet.

Reasoning & sources

If following this BNC campaign is a requirement, consider alternatives to paid Google / Alphabet products where practical. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well. This label follows the named company’s campaign status; it does not claim that BDS separately reviewed this exact product.

This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.

Israel’s procurement administration says Google was selected with AWS for Nimbus public-cloud services and lists the Ministry of Defense and IDF among project partners. The Guardian’s investigation of leaked agreement documents reports that Nimbus terms limited Google’s ability to suspend service, including after alleged policy violations. Who Profits and AFSC describe cloud and AI use by military/security bodies, including surveillance-related capabilities; these are attributed research claims, not proof of Google’s role in a specific attack. Both companies denied circumventing legal obligations in response to the Guardian investigation.

How Gemma family is used in this relationship is not established. The government record establishes the contract and project participants; the investigative reporting describes contract controls, and research organizations make broader use claims. The reviewed evidence does not establish Google’s contribution to a particular operation, injury, or other specific harm.

All sources