DeepSeek R1 and V3: Model Innovations, Industry Impact, and Future Trends

On January 20, 2025 DeepSeek announced the open‑source release of its latest inference model DeepSeek R1, built on the earlier DeepSeek V3 and trained with reinforcement learning to excel at complex logical reasoning, mathematics, code, and natural‑language tasks while reducing inference cost to about 2% of OpenAI's o1.

DeepSeek V3 is a 671‑billion‑parameter Mixture‑of‑Experts model that activates 37 billion parameters per inference, featuring Multi‑Head Latent Attention (MLA), a novel MoE design, and efficient training that matches leading closed‑source models, especially in Chinese factual knowledge.

DeepSeek R1 focuses on deep problem solving and logical reasoning, introducing rule‑driven reinforcement learning, a two‑stage training strategy, and self‑evolution capabilities; its weights, training code, and data pipeline are fully open‑source, fostering community collaboration.

Expert Opinions

Yann LeCun (Meta Chief Scientist) : Open‑source R1 is a major innovation driving global tech progress.

Satya Nadella (Microsoft CEO) : Praises R1’s open‑source strategy and outstanding inference performance.

Sam Altman (OpenAI CEO) : Impressed by R1’s pricing and performance, noting the competitive pressure it adds.

David Holz (Midjourney Founder) : Highlights the advantage of Chinese data for philosophical and humanities queries.

Dario Amodei (Anthropic CEO) : Views DeepSeek’s progress as part of the expected AI cost‑reduction trend.

Other critics : Question the novelty of the technology and raise concerns about data sourcing and geopolitical implications.

Industry Trends

With R1’s MIT‑licensed weights, domestic chip manufacturers (Huawei Cloud, Chengdu Huawi, Moore Threads, Loongson) are integrating the model into edge and server hardware, while cloud providers (Tencent Cloud, Alibaba Cloud, Baidu Cloud, QingCloud, UCloud) offer free or discounted R1/V3 APIs, driving down inference prices.

Application sectors such as smart automotive, finance, AI‑assisted programming tools, and government services are rapidly adopting DeepSeek, exemplified by deployments at BYD, major banks, and the Hangzhou HR bureau.

Future Model Evolution

Research from Fei‑Fei Li’s team demonstrates a distilled model (s1) that matches the performance of OpenAI o1 and DeepSeek R1 using knowledge distillation and supervised fine‑tuning, trained on 16 NVIDIA H100 GPUs in 26 minutes for under $50 in cloud costs.

Test‑time scaling, combined with high‑quality data and open‑source distillation, is identified as a promising paradigm for next‑generation large‑model applications.

Conclusion

DeepSeek’s unconventional approach—favoring result‑based reinforcement learning over step‑wise supervision and openly sharing model assets—has produced a cost‑effective, high‑performance LLM that challenges established players and reshapes the AI landscape.