DeepSeek

DeepSeek-V3

DeepSeek's open-weight Mixture-of-Experts base model (671B parameters, 128K context), released December 2024 and widely self-hosted. Linked findings reflect documented failure-mode classes; per-version attribution is illustrative unless a finding cites this model directly.

Strengths

Strong, cost-efficient general performance from a sparse MoE; long context; open weights enable inspection and self-hosting.

Known weaknesses

Shares the frontier-wide reasoning, arithmetic and hallucination classes; as a base (non-reasoning) model it relies on prompting for multi-step reasoning.

Findings (0)

No findings linked to this model's versions yet.

Versions tracked

deepseek-v3

Cite this

Qlarify Labs. (2026). DeepSeek DeepSeek-V3 — known weaknesses. Retrieved from https://labs.qlarify.fi/models/deepseek-v3