Pleias-Pico
353M-parameter specialist RAG model under Apache 2.0. The developer says it uses permissively licensed Common Corpus data and publishes the training settings.
Score calculation
(7 + 7 + 9) ÷ 3 = 7.7
3 of six dimensions rated. Unrated dimensions and BDS status are excluded.
Full scoring method ↗Setup details
Evidence is from the developer-maintained model card; claims about rights compliance and copyright output limitations are developer assertions, not independent certification.
The model card identifies Pleias-Pico as a specialized 353M model trained on Common Corpus. The page's collection shows the model was initially published in February 2025; no newer Pico checkpoint was verified here. Checked 2026-10-23.
Energy
↑Pleias's model card reports training at Jean-Zay with 16 H100 GPUs using Nanotron, a 90,000-step schedule and one epoch over its RAG dataset. It does not state elapsed accelerator-hours, device/facility electricity, power draw, emissions, or water; steps and GPU count are training-scale context, not energy measurement. No standardized inference-energy figure is reported.
Reasoning & sources
The card gives GPU count and training steps, which describe scale but not elapsed accelerator-hours, facility power, emissions, or water. A comparison needs measured energy and serving data.
No elapsed job duration/GPU-hours, kWh, CO2e, water, or per-task inference energy disclosed.
Training rights
↑PleIAs says Pico is trained only on permissively licensed Common Corpus content, with dataset and training configuration linked. This is a developer claim; no external rights audit is cited on the card.
Reasoning & sources
Pleias says Pico uses permissively licensed Common Corpus material and publishes the dataset and configuration. This is a stronger provenance claim than a generic web-data statement, but it is still developer-reported and should not be treated as a completed independent audit.
Published corpus and rights-handling disclosures are unusually detailed, but remain developer-reported and lack independent rights verification; score reflects transparency with that material limitation.
Independent verification of each underlying license/source is not established.
Privacy
↑Downloadable weights support local inference, so prompt privacy can be controlled by the deployment operator. The card does not establish privacy of hosted third-party RAG applications or endpoints.
Reasoning & sources
Downloadable weights make local inference possible, so a deployment can keep prompts under operator control. The card does not cover hosted providers or local logging; check the actual serving setup before handling sensitive prompts.
Public weights enable local deployment and operator control over prompts, but that is a deployment option rather than a guaranteed privacy safeguard; hosted services and downstream configurations require separate review.
No hosted service retention or telemetry policy assessed.
Openness
↑The model card lists Apache 2.0, public weights, a Common Corpus dataset link, and training settings. This gives users substantial inspection/reuse access, subject to checking the exact dataset and dependency terms.
Reasoning & sources
The Apache 2.0 license, public weights, dataset link, and training settings allow substantial inspection and reuse. The remaining limitation is that source-level rights and actual reproduction still require checking the linked data and environment.
Apache 2.0 weights, data link, and training settings are published.
Verify associated data/license artifacts and dependencies at the chosen revision.
Worker treatment
↑the model card credits compute and technical contributors but does not disclose annotation workers, pay, working conditions, or contractor protections.
Reasoning & sources
The model card credits compute and technical contributors but does not say whether human data annotators were used or how any were compensated. Technical credits do not establish workforce conditions; ask for labor sourcing details.
No labor-supply-chain disclosure found.
Government & military
↑Pleias-Pico is published with downloadable weights under Apache 2.0 and local inference instructions. A downstream user could deploy it without a Pleias service contract; no cited source identifies military adoption.
Reasoning & sources
The model card exposes the weights under Apache 2.0 and gives local inference options. Independent deployment therefore need not appear in Pleias customer disclosures, but the license itself says nothing about actual military use.
No downstream-use register for open weights.
BDS status
↑We did not find PleIAs among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.
Reasoning & sources
PleIAs is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well.
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.