Although DeepSeek V4 was not released during the Spring Festival, DeepSeek released a new model on February 11, called DeepSeek V4 Lite, which has only 200 billion parameters and is relatively small in scale. The main feature of DeepSeek V4 Lite is the 1M context. In previous tests by netizens, its performance was not too impressive except for the ultra-long context. After all, the parameter scale is much different from that of mainstream large models.
However, DeepSeek V4 Lite has been quietly upgraded since its release. On February 27, some netizens tested it and found that its performance is getting more and more powerful. Yesterday, it was upgraded again. HCPTangHY, the leader of the Linux Do community, was also shocked after the test, saying that it was a bit amazing. In his test, it was already the national model SOTA (note: SOTA can simply be considered first).
Not only that, if training continues, it is possible that open source can leverage closed source.
Judging from his tests, the score of DeepSeek V4 Lite's 0302 upgrade is one level higher than the previous 0227 version.It is already close to the current top large model Sonnet 4.6.
There are other games and front-end tests in the post, including the well-known weather card test in the industry. DeepSeek V4 Lite 0302 version is online both in terms of aesthetics and functionality.
Looking back at the large AI models in the past year, the domestic large models are no longer much different from the closed-source models in terms of dialogue and chat.But it is left behind in aspects such as multimodality, programming, mathematics and intelligent agents.The recently released GLM5, MiniMax 2.5, and Qwen 3.5 series have improved, but Anthropic, OpenAI, and Google have also released the latest models, and the gap has widened again.
In catching up with the top closed-source models, domestic companies face many disadvantages. They do not invest as much as American companies, and their computing power is not enough. Data is not as good as Google and OpenAI. After all, they either have years of accumulation or a large number of users.
Moreover, in order to collect data, Anthrpoic downloaded a large number of books from pirated websites and was ordered to pay compensation of US$1.5 billion. This shows how much they invested in this area and how shameless they are.
As the hope of the whole village, DeepSeek has set a benchmark in technology exploration. If DeepSeek V4 Lite with 200 billion small parameters performs like this, then the full version of DeepSeek V4 is indeed worth looking forward to. After its release, it will inevitably have a huge impact on American AI.
