Original | Odaily (@OdailyChina)
Author | Wenser (@wenser 2010)
After Xiaohongshu topped the US App Store due to the Tiktok ban, another "dark horse" Chinese product has topped the App Store in the AI field.
What's quite different is that the significance of this topping is even far beyond Xiaohongshu - according to media reports, the DeepSeek app topped the free APP download ranking charts in the Apple App Store in China and the US, surpassing the "benchmark product in the AI field" ChatGPT. At the same time, the concept Tokens related to the Deepseek R 1 model have also been hotly pursued by the market due to market attention, with the market capitalization of multiple Tokens exceeding the $10 million threshold.
Odaily will briefly introduce the Deepseek product and related concept Tokens in this article for readers' reference.
Deepseek: Challenging OpenAI, a better and cheaper R 1 model than ChatGPT
The origin of all the heat lies in an article released by the Deepseek official 7 days ago.
On January 20, an article titled 《DeepSeek-R 1 Release, Performance Benchmarked against OpenAI GPT-3》 was released, directly challenging the GPT-3 model of AI giant OpenAI, with a detailed introduction to Deepseek's R 1 model, mentioning that:
DeepSeek-R 1 used large-scale reinforcement learning techniques in the post-training stage, and greatly improved the model's inference capability with only a small amount of annotated data. In tasks such as mathematics, coding, and natural language reasoning, the performance is on par with the official version of OpenAI GPT-3.
We will fully disclose the training technology of DeepSeek-R 1 in the hope of promoting full exchange and innovative collaboration in the technical community. Paper link: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf
The pricing of the DeepSeek-R1 API service is 1 RMB per million input Tokens (cache hit) / 4 RMB (cache miss), and 16 RMB per million output Tokens.
In simple terms, not only does Deepseek's R 1 model match the performance of ChatGPT GPT-3, but it also directly open-sources the relevant training technology, and its API pricing is even shockingly low.
Just as many people were still doubting whether Deepseek was "bragging", the subsequent verification results of industry insiders confirmed: Deepseek was telling the truth! And the sharp product that can simultaneously cover the B-end and C-end related needs has also caused countless US AI companies and teams to exclaim: "The wolf is here!"
A PhD student team at the University of California, Berkeley, Jiayi Pan, successfully reproduced the key technology "Aha Moment" in DeepSeek R 1-Zero at a very low cost. The research shows that even small-scale language models can develop strong self-verification and search capabilities through reinforcement learning (RL). Notably, the research team stated that the cost of training the model is less than $30. The aforementioned "aha moment" is a key technology in DeepSeek R 1-Zero.
According to the technical report of DeepSeek-R 1, through the RL framework, AI may spontaneously form human-like reasoning capabilities, even beyond the constraints of pre-set rules. And this may also provide a direction for developing more autonomous and adaptive AI models, such as dynamically adjusting strategies in complex decision-making (medical diagnosis, algorithm design).
Previously, a Meta employee posted on the anonymous US workplace community TeamBlind, stating that due to the superior performance and extremely low training cost of DeepSeek's models, Meta's engineers are frantically analyzing DeepSeek and trying to copy part of its content.
In addition, AI blogger scaling 01 comparative analysis shows that not only has DeepSeek achieved surpassing performance, but the training cost of its 671B parameter DeepSeek-V3 model is only $5.58 million. In comparison, the computational cost of Meta's Llama 3 series models is enough to train DeepSeek-V3 at least 15 times.
—— Excerpt from APPSO related article
It is worth mentioning that the behind Deepseek is the well-known domestic quantitative fund Fantom, which previously hoarded a large number of high-performance graphics cards for quantitative trading needs; and because it belongs to a financial company, it has not been included in the US technology company sanctions list, and finally, relying on its extremely strong determination and long-term technical route, it has achieved the current "using smaller data volume to achieve better output effect" results.
Of course, the Meme Tokens, which have always been the weather vane of news hotspots, are still leading the market.
Deepseek-related Concept Tokens: A market hotspot where technology and Meme intersect
The following are the Deepseek-related concept Tokens that have received some attention in the current market.
seek: The namesake AI concept Token, with a market cap of about $22.7 million
As early as last December, when the Deepseek V3 version was released, many people noticed its speculative value, and seek is the earliest "riding" the Deepseek concept namesake AI Meme Token, which was born 20 days ago.
According to GMGN data, its current price is reported at $0.022, with a market cap of about $20.6 million. Its market cap once exceeded $33 million, and the 24-hour trading volume is about $54.6 million. It is worth mentioning that the Deepseek official X platform account had previously posted that Deepseek official has not issued any Tokens.
seek Token information
POLAR: An AI Copilot built on Deepseek, with a market cap of about $11 million
Polar's main concept is "the first AI copilot on the Solana ecosystem, built on the Deepseek framework", and according to its latest tweet, it has achieved a series of milestones in just 3 days since its launch, and will soon open the Polar API.
According to GMGN data, its current price is reported at $0.011, with a market cap of around $11 million. Its market cap once exceeded $22 million, and the 24-hour trading volume is about $36.4 million.
POLAR Token information
AAA: Launched by a former Amazon programmer, with a market cap of about $6.1 million
AAA is a concept Token related to the AI Agent launched by a former Amazon programmer nim, and according to its tweet information, it uses the Deepseek model and the cost is directly reduced by 98% compared to the ChatGPT GPT-3 model.
According to GMGN data, its current price is reported at $0.0061, with a market cap of around $6.1 million. Its market cap once exceeded $7.5 million, and the 24-hour trading volume is about $4.5 million.
AAA Token information
Latest news shows that a16z founder Shaw posted that "By the way, Eliza has completed support for DeepSeek two weeks ago." It has to be said that a16z is still at the forefront of tracking the AI model track. In the future, we will also continue to follow more AI-related concept Tokens.
Here is the English translation:In addition, Odaily reminds you: The price of MEME tokens fluctuates greatly, and the above content does not constitute investment advice. Please participate cautiously and DYOR.
