What DeepSeek-V3's Cost Story Really Proves
DeepSeek-V3 is a test of whether a model can be simultaneously very large in stored capacity, selective in per-token computation, and economical to train without sacrificing useful capability. The final question is not whether 671 billion parameters are “cheap” in the abstract. It is what the report actually establishes when architecture, routing, memory, numerical precision, training data, evaluation, and deployment are considered together. That judgment matters because the same design that reduces one bottleneck can move cost into another.
The evidence supports a strong but conditional conclusion. DeepSeek-V3 demonstrates a coherent form of algorithm–software–hardware co-design: sparse experts reduce active feed-forward work, MLA reduces persistent attention state, routing control and DualPipe make distributed execution more viable, selective FP8 reduces repeated arithmetic and data movement, and later training stages turn the base model into a usable assistant. The report's own numbers also show the boundary of the achievement. The official training account excludes earlier research, serving requires a large topology-aware cluster, communication still consumes compute resources, and longer reasoning answers can spend more tokens.
Capability is real, but the comparison has conditions
The capability evidence is strong enough to make the efficiency question consequential. In the report's chat comparison, DeepSeek-V3 scores 88.5 on MMLU, 75.9 on MMLU-Pro, and 59.1 on GPQA-Diamond. Those measures cover broad knowledge, harder educational questions, and expert-level questions, respectively. It scores 91.6 F1 on DROP, a reading and reasoning task, and 90.2 on MATH-500. On code-oriented measures, it reports 82.6 on HumanEval-Mul, 40.5 on LiveCodeBench with chain-of-thought, 51.6 on the Codeforces percentile measure, and 42.0 resolved issues on SWE-Bench Verified. The pattern is more informative than one headline score: the model is competitive across knowledge, mathematics, algorithmic code, and software-engineering tasks.
The model's strengths are not uniform. The report says DeepSeek-V3 leads the listed open-source systems on many measures and is competitive with closed models, but it trails Claude-3.5-Sonnet on some engineering tasks and on GPQA-Diamond. Its English SimpleQA score is 24.9, below GPT-4o and Claude in the comparison, while its Chinese SimpleQA score is 64.8 and exceeds the listed alternatives. The authors connect that difference to the training allocation, which gives substantial attention to Chinese knowledge. This is a useful example of capability being shaped by the data and objectives rather than appearing as one undifferentiated property called intelligence.
The long-context results reinforce the same point. DeepSeek-V3 reports 73.3 on FRAMES and 48.7 on LongBench v2, and the report describes robust performance on Needle In A Haystack tests up to 128K after the extension and supervised fine-tuning stages. These results support a meaningful long-context capability under the stated tests. They do not show that every 128K-token input receives equally strong reasoning, retrieval, or factual treatment. A targeted needle test and a benchmark over long documents answer narrower questions than production use over arbitrary lengthy material.
Open-ended evaluations add another layer of evidence. DeepSeek-V3 reports 85.5 on Arena-Hard and 70.0 on the length-controlled version of AlpacaEval 2.0. On RewardBench it reports an average of 87.0, increasing to 89.6 with majority voting across six evaluations. These measurements suggest strong judged interaction quality and judgment ability, but they depend on the evaluation protocol. Arena-Hard and AlpacaEval use an LLM judge, and the report identifies GPT-4-Turbo-1106 as the judge for those pairwise comparisons. A judge score is evidence about performance under a judging system, not a direct measurement free from evaluator preferences.
The comparison settings deserve to remain attached to the numbers. The authors use their internal HAI-LLM evaluation framework, call closed models through their APIs, use benchmark-specific prompts, and limit outputs to 8,192 tokens.