ERNIE 4.5 family
Multilingual text and vision-language models; the 21B/3B-active reasoning checkpoint has a published Apache 2.0 license and direct local inference instructions.
Score calculation
(2 + 3 + 7 + 1 + 7) ÷ 5 = 4.0
5 of six dimensions rated. Unrated dimensions are excluded from the average.
Full scoring method ↗Setup details
Released ERNIE 4.5 text and vision-language checkpoints; excludes ERNIE-Image and later unrelated portfolios.
What fits your needs
Released ERNIE 4.5 text and vision-language checkpoints; excludes ERNIE-Image and later unrelated portfolios.
Good fit
Multilingual text and vision-language models; the 21B/3B-active reasoning checkpoint has a published Apache 2.0 license and direct local inference instructions.
Tradeoffs to expect
Tradeoffs
Apache 2.0, as stated on the official ERNIE 4.5 card. The official card provides downloadable weights, a license, code, and local inference instructions. The reviewed source-level training detail remains limited, so reproducibility and training data transparency are partial.
How to switch
Local setup
The 21B/3B-active card supports local inference, but total checkpoint size and activation memory are material. Check deployment hardware and exact runtime for each dense or MoE variant.
Practical questions
Weights and license
Fetch the exact checkpoint revision and read its own license; sibling and distilled variants may inherit different terms.
Hardware
The 21B/3B-active card supports local inference, but total checkpoint size and activation memory are material. Check deployment hardware and exact runtime for each dense or MoE variant.
Training data
Treat the weight license and public web/data availability separately from evidence of permission for the training data.
This is a practices and disclosure score for the named release, not a request-level energy or emissions estimate.
See the evidence
The official card provides a self-hosted runtime path but no measured electricity, emissions, or water. This score rewards a stated operating boundary; it does not infer impact from missing measurements.
1 + 1 reporting + 0 electricity + 0 controls = 2/10. This grades documented practices and transparency.
No per-request electricity, emissions, water, serving mix, or independently reviewed footprint is established in the cited sources.
- Reporting1/3
- The official release materials provide downloadable weights and a local serving path for ERNIE 4.5; this establishes an operating boundary, not a measured footprint.
- Electricity sourcing0/3
- No model-specific clean-electricity sourcing or hourly power evidence is published in the reviewed sources; zero means unestablished, not dirty electricity.
- Resource controls0/3
- No model-specific documented control or measured physical resource reduction is established by the reviewed release materials. Public weights or smaller variants alone do not prove reduced resource use.
Water disclosure
Water reporting has not been assessed for this choice.
The release provides some training-process information, but not a complete permission ledger for inherited and later training data.
See the evidence
Baidu publishes model and technical documentation, but the reviewed sources provide limited named training-source detail and no complete training data permission or creator-remedy audit. A public weight license governs use of those weights; it is not proof that every training source was permitted.
1 base + 1 source transparency + 1 documented permission + 0 creator controls = 3/10. Weight licensing is not treated as training data permission.
Complete source lists, source-specific rights, inherited-base provenance, synthetic-data generation permissions, and creator removal or retraining remedies remain unestablished.
- Source transparency1/3
- Baidu publishes model and technical documentation, but the reviewed sources provide limited named training-source detail and no complete training data permission or creator-remedy audit.
- Documented permission1/4
- Some named or permissively licensed material is described, but the sources don’t settle permission for the full inherited and later-stage training data.
- Creator controls0/2
- The reviewed model documentation doesn’t show creator permission covering all training data, opt-out, removal, or retraining remedy for this model family.
Privacy depends on where these released weights are run; the model family has no single prompt-retention or training route.
See the evidence
The official release makes weights available for separate deployments, but it does not define one inference host or retention policy for every user. Local and hosted operators may handle prompts differently, so a single numerical privacy grade would misstate this model-level entry.
The user-selected host, region, telemetry, logging, extensions, and provider training terms are outside the checkpoint release and require a route-specific review.
The official card provides downloadable weights, a license, code, and local inference instructions. The reviewed source-level training detail remains limited, so reproducibility and training data transparency are partial.
See the evidence
The official card provides downloadable weights, a license, code, and local inference instructions. The reviewed source-level training detail remains limited, so reproducibility and training data transparency are partial. This openness rating concerns artifacts and reproducibility only; it does not assess whether the original training data was authorized.
Apache 2.0, as stated on the official ERNIE 4.5 card. Model weights and documentation are available, but limitations described above reduce source-level reproducibility.
The checked release and license apply to the named checkpoints; sibling variants, future versions, and the underlying training data may have different terms or disclosures.
No model-specific worker safeguards are established for Baidu in the reviewed model sources.
See the evidence
The model release pages describe weights and technical properties, not conditions for employees, data annotators, evaluators, moderators, or infrastructure contractors. This is a source-bounded disclosure gap, not evidence of mistreatment.
1 base + 0 conditions + 0 worker voice/remedy + 0 evidence/follow-through = 1/10; this is a disclosure finding, not proof of abuse.
Worker locations, pay and conditions, supplier coverage, voice mechanisms, labor audits, and remediation outcomes are not established by these model-release sources.
- Pay & conditions0/3
- No worker-level conditions or supply-chain safeguards are established in this model-family source set. This is a disclosure gap, not proof of mistreatment.
- Worker voice & remedy0/3
- The model sources do not document a worker voice or remedy mechanism for data, evaluation, or infrastructure labor.
- Evidence & follow-through0/3
- No independent worker-level audit or remediation outcome is established in the reviewed model sources.
No military relationship or release-specific safeguards are established for Baidu in the reviewed model sources.
See the evidence
The official model card and release materials identify the model and license but do not provide a military customer, contract, deployment, or model-specific enforcement record. A relationship may still exist even if this review found no record of it.
7 base + 0 safeguard adjustment = 7/10 (0 of 9 safeguard points; every 3 adds 1). No military relationship was established in the cited sources. This ranks above known defense ties without claiming a verified clean record.
Undisclosed contracts, customer-specific deployment, end-use restrictions, and independent compliance outcomes are not established by this model-family review.
- Use restrictions0/4
- The cited model license and release documents don’t show a model-specific binding ban on military or intelligence use. This is limited to this source set.
- Enforcement & oversight0/3
- No military-specific monitoring, enforcement, or remedy mechanism is established in the reviewed model materials.
- Relationship disclosure0/2
- The model materials do not disclose a model-specific military relationship or military-use record. A relationship may still exist even if this review found no record of it.
The reviewed model-release sources don’t confirm a defense relationship; absence of disclosure is not a clean-record finding.
Military rubric ↗Baidu is not named in the two official BNC campaign lists checked. This is a limited list check, not BDS approval.
See the evidence
The official BNC consumer boycott/pressure guide and No Tech campaign page do not name Baidu among their current targets. BDS uses selective campaigns, so absence from these lists is neutral and should not be read as a clean record or approval.
This check covers only named targets in the two cited BNC campaigns; it does not cover all BDS campaigns, local groups, investments, suppliers, or downstream hosts. Lists and ownership may change.