推文

Why does DeepSeek have more influence than Qwen when their performance is so close? 🌟Insights from Zhihu contributor 情酱 Here’s a view that may not be universally popular: DeepSeek’s bigger influence is not really about which model is stronger. The real gap comes from something else: DeepSeek is willing to show the entire research process. 🔍 I’ve been following both DeepSeek and Qwen’s technical reports since the V3 era. The difference is obvious. Reading DeepSeek’s papers feels like sitting down with an engineer who walks you through the whole journey: what they tried, where they failed, why they chose one path over another, and even which experiments did not work. A recent example is DeepSeek V4, released on April 24. Its 58-page technical report was uploaded to Hugging Face on the same day. Chinese tech media even described it as “surprisingly detailed.” Before V4, DeepSeek released several full research papers, including mHC, Engram, and DualPath. When the V4 report came out, the community could check directly: which techniques landed, which did not, and why. But the part I respect most is not the technical detail itself. It is the honesty. 🧠 In the V4 report, DeepSeek essentially said that V4-Pro-Max outperforms GPT-5.2 and Gemini-3.0-Pro in reasoning, but still slightly lags behind GPT-5.4 and Gemini-3.1-Pro. It also admitted that its development trajectory is roughly 3 to 6 months behind the frontier closed-source models. Think about that. An AI company wrote in its own technical report: “We are still 3 to 6 months behind the very best models.” That level of transparency is rare in the AI industry. R1 followed the same philosophy. The paper explained how GRPO was designed, why PPO was not used, and listed hyperparameters and learning rates. It even included a Negative Results section, explaining that PRM and MCTS were tried and failed, along with the reasons behind those failures. Later versions added more training details, including the RL training architecture based on vLLM. Why does this matter? Because research teams around the world can actually reproduce the work. 🌍 Hugging Face’s Open R1 team trained models using DeepSeek’s public methods, and the results were close to the official distilled version. That is much more persuasive than posting 100 benchmark charts. Now compare this with Qwen. To be clear, this is not about attacking Qwen. Qwen’s engineering capability is genuinely impressive. Over the past few years, it has open-sourced hundreds of models. The Qwen3.5 series covers everything from 0.8B to 397B parameters. It uses Gated Delta Networks, Early Fusion for native multimodality, a Thinker-Talker architecture for end-to-end audio and video, and a 397B-parameter model that activates only 17B parameters. The technical depth is real. ⚙️ But when you read Qwen’s technical reports, the feeling is different. For Qwen2.5, some readers said the report could be finished in 10 minutes and did not go deep into the technical principles. Much of it focused on benchmark comparisons. With Qwen3 and Qwen3.5, even though the architecture changed meaningfully, it is still hard to find much discussion about training problems, failed attempts, abandoned directions, or lessons learned. That creates a subtle contrast. Every time Qwen releases a new model, the marketing noise is huge: “beats DeepSeek,” “far ahead,” “the world’s No.1 open-source model.” But once the hype cools down and people read the report, there often isn’t that much new methodology to learn from. And those benchmark gaps can disappear in the next iteration. Over time, people start to feel: big thunder, light rain. DeepSeek is almost the opposite. Its releases are not always loud. V4 took 484 days to arrive, with constant rumors and speculation in between. But once the paper dropped, the 58-page report explained everything clearly, from mHC to hybrid attention to the Muon optimizer. The community reaction was immediate. 🚀 To me, these two companies represent two very different open-source philosophies. Qwen feels more like a big tech product route: “I give you the model weights. You can use them. My report tells you how strong I am and where I rank on benchmarks.” That is open-sourcing the result. DeepSeek feels more like a research-driven route: “I give you the model, but I also show you the full research process: the training pipeline, failed paths, hyperparameters, trade-offs, and even how far I still am from the best closed-source models.” That is open-sourcing the process. 🔬 And for the community, the difference is huge. If you give me model weights, I can fine-tune, deploy, and build products. But if you give me a complete training methodology, I can stand on your shoulders, improve your approach, and explore directions you have not tried yet. GRPO is the best example. It was first released with DeepSeekMath in April 2024 and did not attract much attention at the time. But after R1, the entire industry rushed to reproduce it. Kimi, GLM, and many other teams started using similar ideas. By V4, DeepSeek even borrowed the Muon optimizer from Kimi to replace AdamW, and openly credited the source in its paper. That is what a healthy open-source ecosystem looks like: 🔁 • Your method can be reproduced by others. • My method can be adopted by you. • Everyone pushes the field forward together. Qwen is also open-source, but more often it shows the moves rather than teaching people how to train. So why does DeepSeek have greater influence? Not because it crushes Qwen on performance. In many benchmarks, the two are actually close, and Qwen is stronger in some areas. DeepSeek V4 itself admits it still trails the top closed-source models by 3 to 6 months. The real reason is that DeepSeek did something harder: it publicly exposed its failures, limitations, and full methodology. A team that dares to write “we are still 3 to 6 months behind” in its own paper is hard not to trust. Influence is not built by chasing benchmark scores. It is built through trust, one paper at a time. ✍️ 🔗Original article： https://t.co/aBs1aXb8j3 #DeepSeek #Qwen #OpenSourceAI #LLM #AIResearch #MachineLearning #AI