The young team "small steps and fast running" first revealed their ambition: not to use application gimmicks, and to fight against the bottom of AGI!
Tonight, a thunder thunder sounded from the Chinese mockup battlefield!
Just as the world is closely watching OpenAI and Google's new product launch conference, China's AI dark horse DeepSeek suddenly opened source R1's latest 0528 version - no preview or momentum, a series of "We Open Source" announcements made the domestic and foreign developer communities boil in an instant!
Hard-core test: Multiple indicators tie GPT-4
Authoritative tests show that the 0528 version has achieved "fault evolution" in key areas:
- Mathematics ability soared by 87% (GSM8K test reached 89.5%), high school students have difficulty math problems;
- Code generation accuracy crushes Claude 2 (HumanEval score 81.2%), Python/C++ dual cultivation is no pressure;
- 100,000-word long article reasoning zero omissions, contract review and paper study error rates are close to zero!
(Netizens exclaimed: "The Chinese scene has surpassed GPT-4!")
Behind the small steps and fast running: the "technical tortoise and hare race" of the Chinese team
When the industry gathers together to package AI applications to make quick money, DeepSeek chooses a "stupid road":
> ▸ Try to the essence of reasoning - give up commercial gimmicks and focus on improving the depth of the thinking chain;
> ▸ Open source is a weapon - open the latest achievements for free and build a 10,000-person developer alliance;
> ▸ Young legion raid - Tsinghua University and Peking University + Silicon Valley background core team, average age 28 years old!
"AGI is not a press conference PPT," the team's internal letter revealed: "We need to be a probe that penetrates the deep waters of the technology."
Industry shock: open source ecology may be reshuffled
Industry insiders urgently interpret DeepSeek's strategic killer move:
- The country's computing power is independent - the entire process is adapted to domestic chips, breaking the deadlock "Nvidia dependence";
- Defining China's AGI standards - using open source to force the industry to return to the essence of technology;
- Create a "golden entrance" for developers - call top models at low thresholds, and small and medium-sized companies are about to make a counterattack!
The golden sentence at the end of the article:
> "Silicon Valley has closed source building, and China has built a bridge with open source building -
> What DeepSeek tear apart is not the version number, but another possibility of AGI."
Author's Note: The technical parameters and strategic analysis of this article are based on the following authoritative information
1. DeepSeek official open source announcement and technical report (GitHub:
deepseek-ai/deepseek-R1-0528)
2. The latest list of the global big model evaluation platform OpenCompass (2025.06)
3. "Big Model Open Source Ecology White Paper" by the Chinese Society of Artificial Intelligence
4. Interpretation of industry experts' interviews and team technical paths (with official authorization from DeepSeek)
