DeepSeek's new model is online: DeepSeek-V3 with 685 billion parameters has evolved further!
The new DeepSeek-V3 model with 685 billion parameters is here, a double leap in performance and stability!
Core content:
1. The number of parameters of the DeepSeek-V3-0324 model is the same as the previous generation, both of which are 685 billion. 2.
Supports three different precision floating point formats: BF16, F8_E4M3 and F32.
3. Performance improvement and bug fixes, the two major improvements of DeepSeek-V3-0324
DeepSeek finally released a new model!
On the evening of March 24, DeepSeek uploaded a video called DeepSeek-V3-0324 New model.
As the name suggests, this new model from DeepSeek is a successor to the previous generation DeepSeek-V3 An upgraded version of .
Click to enter DeepSeek-V3-0324 Although DeepSeek has not yet uploaded a detailed README document, we can still infer some clues from the model parameter information on the right.
The new model has 685B parameters, or 685 billion parameters. This number is similar to the previous generation. DeepSeek-V3 It's the same.
The tensor data types supported by the new model are also consistent with the previous model: it supports three different precision floating point formats: BF16, F8_E4M3 and F32 for inference and training calculations.
There is no doubt that DeepSeek, which has been committed to open source, has launched DeepSeek-V3-0324 It is also a completely open source model.
Although DeepSeek has not officially introduced it, it is not difficult to guess that this time DeepSeek-V3-0324 The improvement will mainly be in two aspects.
One is performance. DeepSeek-V3 To be honest, it is quite powerful. My WeChat public account is currently connected to DeepSeek-V3 Model.
And I am also DeepSeek-V3 I tested it when it was first released: After testing DeepSeek V3 and GPT-4o, I don’t want to open a ChatGPT membership anymore! This review article was published before DeepSeek became popular, and there were still many cute people questioning it at that time. DeepSeek-V3 ability.
The second is to fix bugs. Yes, you read that right, the model also has bugs. DeepSeek-V3 There is a very "fatal" problem: function call loop and empty response.
This is the announcement that DeepSeek officially released on its open platform. But now, this sentence is gone.
Where to use DeepSeek-V3-0324, DeepSeek officials have not yet made it clear. But Hugging Face has already said that the underlying model in the "non-deep thinking" mode of the official website has been replaced with DeepSeek-V3-0324.
In fact, there is no need to rush.DeepSeek-V3-0324 Once officially released, the DeepSeek official website and API will be updated synchronously.
