Deepseek still drops a small bomb with the update of his V3 model

It is without drum or trumpet that Deepseek has discreetly updated its V3 language model. A fairly amazing approach when you remember the media noise generated recently by the presentation of the Deepseek R1 reasoning model. Recall that the latter offered the advantage of being much more affordable for use than its Western equivalents.

Deepseek returns this time with an important update of his LLM V3. The latter had already been praised for his ability to compete GPT-4O Or Geminiwhile displaying costs that are still clearly lower. The Chinese startup has put online an updated version Deepseek-V3-0324. As said above, everything was in silence, the company having published no press release or blog post to support this launch. This silence as astonishing as voluntary recalls the practices of Mistral, another player in the sector who favors minimal communication.

The same architecture, in addition massive

On the technical level, Deepseek-V3-0324 is based on an architecture mixture-of-experts (MOE), which activates only part of the parameters with each iteration. This operation makes it possible to limit material needs while maintaining a high calculation capacity. The new model now includes 685 billion parameters, against 671 billion for the previous version, an increase of 14 billion.

The total “weight” is around 700 GB, which makes its use possible on certain high -end work stations, such as the most powerful Mac studio configurations. The feedback from the first users report tangible progress. In particular, they report better performance for the resolution of mathematical tasks, with larger differences than expected compared to the previous version.

Always open source, and accessible via Hugging Face

As for its previous models, Deepseek offers the V3-0324 model for free download on Hugging Face. Users who have a machine capable of collecting the beast can thus experience the model locally, without depending on an API or a cloud service. An obviously fundamental difference compared to Openai, which maintains its proprietary models under strict control. This strategy is for the moment paid, because it has already allowed Deepseek to arouse the interest of many actors, including outside China.

Development costs always under control

One of the peculiarities of Deepseek models is due to their production cost, from what the startup says. The training cost of this new version of V3 has not been announced, but would remain lower than that of GPT-4O by a factor ten according to several estimates, which therefore remain to be manipulated with tweezers. But even if it turns out to be a little less in the end, such a gap will again raise questions about the prices charged by the American laying of AI, such as Openai or Google. However, in the absence of official documentation detailing the performance of Deepseek-V3-0324 for the moment, it remains difficult to precisely assess its positioning compared to GPT-4O or Gemini 2.0. It will be noted all the same that the first returns seem to point to a very good level of overall performance, which is enough to make it a serious competitor. And above all, we must not forget that Deepseek must also keep the models of his very active compatriots to the eye, such as Alibaba and his Wan,, Baidu and his Ernieor Manusin another register.