Ethical Stack

DeepSeek V4 family

The latest verified API release is DeepSeek V4.1 Flash, updated October 17, 2026; DeepSeek V4 Pro remains available per the changelog.

Score calculation

(4 + 4 + 8) ÷ 3 = 5.3

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

Full scoring method ↗
Setup details

Release availability is from the official API changelog. The requested social dimensions are only assessed where a primary disclosure was located; absence of public detail is recorded as a gap.

Official API changelog says V4.1 Flash was upgraded to Flash-1017 on October 17, 2026 and V4 Pro would continue to be served. Checked 2026-10-23.

Energy

↑

no V4.1 model-specific energy, water, or emissions data found. Efficiency and pricing claims cannot stand in for energy measurement.

Reasoning & sources

The API update log documents product versions but does not quantify V4.1 training or inference energy. Capability and pricing claims cannot fill that gap; a comparison needs measured training compute and serving data.

No training compute or inference energy report in checked first-party material.

Training rights

↑

DeepSeek’s public V4 training-content summary (last updated April 8, 2026, and identifying V4-Pro/V4-Flash) reports more than 10 trillion text tokens; names Common Crawl, Stack Exchange, broad public corpora, commercially licensed data, and non-public third-party datasets; and describes DeepSeekBot collection with robots.txt/rights-reservation measures. It also marks user-interaction data and synthetic data as used, with de-identification and opt-out language. The summary does not identify every source or establish that all these details apply to the later V4.1-Flash-1017 update.

Reasoning & sources

DeepSeek’s V4 summary names major source categories and datasets, commercial licensing, a crawler that respects rights reservations, and user-data opt-out provisions. Because the report identifies V4-Pro/V4-Flash and predates V4.1-Flash-1017, it is useful family context rather than proof of the newer checkpoint’s complete provenance.

The V4 disclosure identifies source categories, named corpora, licensing, and crawl controls, but exact V4.1-Flash-1017 coverage and independent rights clearance are not shown.

The April 2026 V4 summary is not a V4.1-Flash-1017-specific data ledger; exact update lineage, dataset-level licensing, and independent rights verification remain unavailable.

Privacy

↑

The official Responses API reference says responses and conversations are not stored on that endpoint; the October 17 changelog maps `deepseek-flash` to V4.1-Flash-1017. This is an endpoint-specific no-conversation-storage statement, not a promise that no operational telemetry or other API data are retained. DeepSeek’s separate first-party app policy describes collection and model-improvement use with opt-out, and excludes downstream developer-run apps.

Reasoning & sources

DeepSeek documents no server-side response or conversation storage for its Responses endpoint, and its changelog connects `deepseek-flash` to V4.1-Flash-1017. That is a narrow retention safeguard, not a general no-logging promise; verify telemetry and other endpoint terms before sending sensitive prompts.

The specific endpoint promises no server-stored conversations, while the broader first-party app policy permits collection and model-improvement use and does not settle API telemetry.

The stateless guarantee covers response/conversation storage for the documented endpoint; API telemetry, other API endpoints, and app-service handling require separate policy checks.

Openness

↑

DeepSeek-V4.1-Flash-1017 is publicly downloadable from DeepSeek’s official model repository under the MIT License, with inference instructions and supporting code. The weights can be run locally, but the complete training corpus and end-to-end training pipeline are not fully public.

Reasoning & sources

DeepSeek released V4.1-Flash-1017 weights under MIT and supplies local inference instructions, so users can download, inspect, and deploy the checkpoint themselves. Training data and the full training run remain less transparent, making this strong weight openness rather than complete reproducibility.

Official MIT-licensed weights and inference artifacts enable broad local use and inspection; the undisclosed full corpus and training pipeline limit reproducibility.

The release provides weights and inference materials but not a complete source-level corpus or end-to-end reproducible training process.

Worker treatment

↑

DeepSeek’s V4 training-content summary says an internal team of professionals creates, labels, and audits data in some domains; the general training-method disclosure says its research team produces manually or automatically prepared optimization examples. Neither source identifies the exact labor behind V4.1-Flash-1017 or reports headcount, pay, employment terms, or working conditions.

Reasoning & sources

DeepSeek reports professional staff creating, labeling, and auditing some V4 data, confirming that human work contributes to the family. The sources do not identify the V4.1-Flash-1017 workforce or its terms, so the ethical quality of that labor remains unclear.

No worker counts, compensation, employment conditions, vendor information, or confirmation of labor practices specific to V4.1-Flash-1017.

Government & military

↑

DeepSeek publishes V4.1 Flash weights under MIT and documents local inference. That makes independent downstream deployments possible without a DeepSeek service contract; it is not evidence that a military customer has deployed them.

Reasoning & sources

The exact checkpoint is MIT-licensed and its card gives local inference instructions. A downstream defense user could therefore run it without appearing in DeepSeek’s hosted-customer list; no source here identifies an actual military deployment, so that relationship remains unverified.

Requires procurement and company source review.

BDS status

↑

We did not find DeepSeek among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.

Reasoning & sources

DeepSeek 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.

All sources

Side by side