New open-source model from DeepSeek leads non-reasoning AI, excelling in real-time applications despite heavy computational demands.
Chinese artificial intelligence (AI) company DeepSeek released an updated version of its V3 model, named DeepSeek-V3-0324. This launch occurred quietly on Hugging Face, a popular AI development platform. Many news outlet including HPCWire, Tom’s Hardware, and ZDNET cover the Deepseek update.
Developers and researchers quickly noticed the new model. DeepSeek presents this upgrade as a refinement of the original V3, launched in December 2024. The new model retains the same mixture-of-experts architecture.
This design enhances efficiency over traditional models. DeepSeek-V3-0324 operates under the MIT license, making it fully open-source. This shift from the previous custom license broadens access for developers worldwide.
The V3 update climbed seven points on the Artificial Analysis Intelligence Index, AINews reports. It now tops proprietary non-reasoning models like Google’s Gemini 2.0 Pro. The upgrade also surpasses Anthropic’s Claude 3.7 Sonnet and Meta’s Llama 3.3 70B. This marks a breakthrough for open-source AI.
Unlike reasoning models, V3-0324 delivers instant answers without “thinking” delays. This suits it for chatbots and live translation needs. It trails behind reasoning models like DeepSeek’s R1 and OpenAI’s offerings. Still, its edge in latency-sensitive tasks stands out. AINews notes it’s the first open-weights model to lead non-reasoning categories.
The model keeps specs from its December 2024 version. It features a 128k context window, capped at 64k via API. With 671 billion parameters, it demands over 700GB of GPU memory. Only 37 billion parameters activate per task. It handles text only, lacking multimodal features.
Impact on AI landscape
Three months ago, V3 nearly matched top proprietary models but didn’t overtake them. Now, V3-0324 leads all non-reasoning rivals, open-source or not. AI benchmarking & analysis company Artificial Analysis finds this release more striking than DeepSeek’s R1. The gap with reasoning models narrows, though complex tasks still favor the latter. Open-source AI gains ground against closed systems. Developers gain a robust tool in V3-0324. Its computational cost, however, poses challenges. Artificial Analysis credits DeepSeek with pushing non-reasoning model frontiers. The AI sector feels this shift. With R2 nearing, anticipation builds for further advances.
Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.
0 Comments
0 thoughts on “DeepSeek unveils upgraded V3-0324 model”