There is about a week left in this month, and the biggest expectation is DeepSeek V4. From various indications, it is getting closer, and the official has been making continuous moves recently. Now it is found that the API has also begun to upgrade the model, and the performance is not bad. Since the announcement of a beta version on February 8, DeepSeek V4 has actually been undergoing low-key changes, but the official has not introduced the upgrade content in detail each time.
Not long ago, the DeepSeek V4 page was upgraded, with three modes available: fast, expert and visual. However, the previous upgrade was mainly for the web version, and now it is found that the model on the API side has also been switched.
I'm just not sure about the specific name of this model.According to the analysis, the DeepSeek V4 Lite with small parameters is the most likely.
Of course, there are still surprises. A leader in the Linux.do community mentioned the information about the internal beta, pointing out that DeepSeek will have two models, which can adjust the thinking gear, support FP8, and have strong command compliance.
What's more,It is mentioned here that its performance is close to Opus 4.5, but complex tasks are not as good as GPT.
Although Anthrpoic has released Opus 4.6 and 4.7, these two large models have been controversial recently. Issues such as intelligence reduction and large increase in usage have directly triggered community backlash. If the larger model in DeepSeek V4 can reach the level of Opus 4.5, it will definitely be the leader among open source large models. This capability is enough for most developers to use, not to mention that it is domestic open source and has many fewer restrictions than A.
But everyone must also be prepared. As the performance of DeepSeek V4 has improved, the cost will not be the same as before, and the model with a large number of parameters will not be cheap.
Although the specific price is not mentioned here, it is definitely higher than DeepSeek V3.