概述
V3 uses a mixture-of-experts architecture; R1 applies reinforcement learning and a multi-stage post-training pipeline to reasoning. Reported GPU-hour figures describe a specific V3 training run, not the full research cost or the cost to train R1.
深入探讨
The DeepSeek-V3 report describes a 671B-parameter mixture-of-experts model with 37B activated per token. It reports 2.788 million H800 GPU-hours for the full V3 training run; this compute estimate is not the total cost of the research program and does not state the cost of training R1. DeepSeek-R1 builds on DeepSeek-V3-Base. Its paper describes a multi-stage pipeline: cold-start supervised fine-tuning, reasoning-focused reinforcement learning, rejection sampling and supervised fine-tuning, then a second RL stage. The paper evaluates verifiable tasks such as mathematics, coding competitions, and STEM. It also describes smaller distilled models based on Qwen and Llama families.
现实世界的实施
Inspecting the V3 technical report’s active-versus-total parameter description for its MoE architecture
Comparing R1’s reasoning evaluation on mathematics, coding, and STEM tasks
Studying the stages used to build R1 from a V3 base model
Using smaller R1-distilled models based on Qwen and Llama checkpoints for research
DeepSeek V3 和 R1 Reasoning 的未来
DeepSeek’s reports contribute open-weight model artifacts and technical descriptions of MoE training and RL-based reasoning. Results should be tied to each model, checkpoint, benchmark, and training stage. Reported compute estimates are narrower than total research, infrastructure, and data costs.
付诸实践
为什么这很重要
- 供应商策略. 供应商路线图会影响您的团队接下来可以构建的功能。
- 成本与预算. 商业条款和部署选项会影响长期成本和风险。
- 风险与安全. 公司激励措施塑造了产品默认、安全态势和开放性。
留意
- 发布公告可能会超过实际生产工作流程的稳定性。
- API 定价或政策转变可能会在一夜之间打破假设。
- 单一供应商依赖性增加了锁定和迁移成本。
后续步骤
- 使用您自己的任务和数据集评估提供商。
- 在集成之前查看隐私、安全和法律条款。
- 维护跨模型或供应商的后备计划。
- 监控发行说明，以便路线图的更改不会让团队感到意外。
不断探索
常见问题
What is DeepSeek V3 and R1 Reasoning?
DeepSeek-V3 and DeepSeek-R1 are open-weight model releases with different technical contributions. V3 uses a mixture-of-experts architecture; R1 applies reinforcement learning and a multi-stage post-training pipeline to reasoning. Reported GPU-hour figures describe a specific V3 training run, not the full research cost or the cost to train R1.
DeepSeek-V3技术报告给出了哪些参数？
The V3 report describes 671B total parameters and 37B active per token in its MoE architecture.
DeepSeek-R1的推理训练和发布有哪些值得注意的地方？
The R1 paper describes reinforcement-learning-based reasoning and the project released model weights; the paper does not imply every capability is achieved by RL alone.
DeepSeek-V3 报告给出的完整训练运行的计算数据是多少？
The V3 report lists 2.788M H800 GPU-hours for the full training run; this is not a full research-cost figure or an estimate for R1 training.
哪些训练信号帮助 DeepSeek-R1 发展推理行为？
The R1 report describes RL using verifiable reward signals for reasoning tasks, alongside other stages in the full R1 pipeline.
与 DeepSeek-V3 训练相关的一种效率技术是什么？
V3 引入了多头潜在注意力和辅助无损失负载平衡方案等技术来有效地训练其 MoE。