DeepSeek-V4-Flash

Item: DeepSeek-V4-Flash
Rating: 2
Author: Ali Madjaji

DeepSeek · China

DeepSeek-V4-Flash is the smaller-active sibling of V4-Pro under the same MIT terms — permissive on the weights, but the China-based vendor and undocumented training corpus mean any EU deployment still needs a self-hosted topology and a deployer-side GPAI documentation file under AI Act Article 53.

Sovereignty

Licence: MITCommercial: UnrestrictedTraining data: Categories onlyOrigin: China

Licence facts

Parameters: 284B total / 13B active
Architecture: Mixture-of-Experts (FP4 + FP8 mixed precision)
Context window: 1M tokens
Training tokens: 32T+
Relation: Smaller-active variant of DeepSeek-V4-Pro

Performance & pricing?

Quality index47/100Speed79 tok/sBlended price$0.17/M

Known risks

Vendor jurisdiction is the People's Republic of China — outside the GDPR adequacy list.
Training corpus published only at aggregate level (32T tokens, broad mix); no dataset card to support AI Act summary-of-content disclosure.
Released as a 'preview' — model card flags ongoing tuning, so behaviour and safety properties may shift between snapshots.

Sources

→ Model Card → LICENSE file