Verdict
DeepSeek-V3 leads 6–1 across 7 shared benchmarks.
DeepSeek-V2 (MoE-236B, May 2024) 1 · DeepSeek-V3 6 · higher isn't always better — check the caveats below.
Reported scores — protocols may differ. How we compare
Commonsense Reasoning: DeepSeek-V3 leads 2–1 across 3 shared commonsense reasoning benchmarks (largest gap: HellaSwag, 85.2 vs 82.8 — single evaluator).
Knowledge: DeepSeek-V3 leads 2–0 across 2 shared knowledge benchmarks (largest gap: MMLU, 82.93 vs 71.2 — single evaluator).
Vendor claims — compare against the measured scores above. A claim is what a developer says about its own model, not an independent measurement.
No vendor claims on file for DeepSeek-V2 (MoE-236B, May 2024).
Coding
“DeepSeek-V3 achieves the best performance on most benchmarks, especially on math and code tasks.”
General
“At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model.”
“DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models.”
Reasoning
“Our pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning performance.”
DeepSeek-V3 leads 6–1 across 7 shared benchmarks. DeepSeek-V2 (MoE-236B, May 2024) leads 1 benchmark and DeepSeek-V3 leads 6. Higher isn't always better — see the integrity caveats on each benchmark.
DeepSeek-V2 (MoE-236B, May 2024) and DeepSeek-V3 have 7 benchmarks in common in our data — those are the only rows where a direct, apples-to-apples comparison is drawn.
DeepSeek-V3 leads knowledge 2–0. DeepSeek-V3 leads reasoning 2–0.