DeepSeek V3.2-Speciale Challenges Google Gemini 3 Pro in Global AI Race
Artificial intelligence start-up DeepSeek has surprised the global tech community with the launch of its most advanced model so far, DeepSeek V3.2-Speciale. The model arrives at a time when the industry is closely watching major developments ahead of the annual NeurIPS conference. Despite facing limits on advanced semiconductor chips, DeepSeek has managed to reach performance levels usually dominated by Google DeepMind and OpenAI.
- DeepSeek V3.2-Speciale Challenges Google Gemini 3 Pro in Global AI Race
The release of V3.2-Speciale has sparked major discussion within the AI research world. Its strong results raise new questions about how much innovation can be achieved even with restricted computing power. Experts have praised the technical report, and the announcement has already begun influencing market reactions and research conversations across the industry.
A Model That Matches Gemini 3 Pro in Key Tasks
DeepSeek said that V3.2-Speciale matches Google DeepMind’s new Gemini 3 Pro in reasoning tasks. Gemini 3 Pro was launched just two weeks earlier, making the comparison even more striking. According to DeepSeek, the base model V3.2, released on the same day, performs on the same level as OpenAI’s GPT 5, which came out in August.
One major highlight is that V3.2-Speciale achieved gold-medal performance on the International Mathematical Olympiad test. This level of achievement has previously been reached only by internal models from OpenAI and Google DeepMind, none of which have been released publicly. This places DeepSeek’s model in an elite category of high-performance AI systems.
Industry Praise and Market Reaction
The launch received positive comments from leading AI researchers. Google DeepMind principal research engineer Susan Zhang praised DeepSeek for the detailed technical report that explained how the team stabilised the model after training and improved its agentic capabilities.
The news also had a financial impact. Shares of Alphabet, Google’s parent company, fell 1.65 percent on Monday after the announcement. This reaction reflects how seriously the market views competition in the frontier model space.
High Performance Despite Lower Compute Access
DeepSeek’s report states that its models required fewer total training FLOPs than competitors in the United States. This is notable because Chinese companies face export controls that limit access to advanced semiconductor chips. FLOPs measure the computing power needed to train large AI models, and having fewer FLOPs usually means less training capacity.
Despite this challenge, DeepSeek managed to produce a model that competes with some of the strongest systems in the world. The company credits this success to training strategies and efficiency improvements rather than raw compute.
However, DeepSeek also admitted that V3.2-Speciale is weaker than Gemini 3 Pro in token efficiency and world knowledge, which comes from having less computational power during training.
Open Source Release and Pricing Strategy
DeepSeek has open-sourced the base V3.2 model on the Hugging Face platform. However, V3.2-Speciale is available only through an API because it uses higher amounts of tokens during operation. The pricing is also significantly lower than competitors.
The model costs 0.28 US dollars per million input tokens and 0.42 US dollars per million output tokens. By comparison, Gemini 3 Pro charges up to 4 US dollars per million input tokens and 18 US dollars per million output tokens. The lower pricing could attract developers and businesses looking for affordable high-performance AI solutions.
Scaling Up to Close the Gap
DeepSeek stated that it plans to improve its future models by increasing the computing power used during training. This approach follows the scaling method used by Google DeepMind, where larger amounts of training data and compute typically result in better models.
Researcher Gou Zhibin wrote that scaling is not hitting any limit. ByteDance’s global head of technology policy, Patrick Zhang, commented that DeepSeek’s achievement shows that compute remains the main factor that decides leadership in the AI industry.
He noted that when models share similar access to knowledge, the teams that train more and scale further usually guide the direction of the field.
Strong Attention Ahead of NeurIPS
DeepSeek’s timing has drawn comparisons to OpenAI’s surprise release of ChatGPT during the 2022 NeurIPS conference. Many experts believe that they planned its release to maximise visibility within the research community. According to attendees, group chats were full of conversation and excitement after the announcement.
This year’s NeurIPS is happening in two locations, San Diego and Mexico City, due to visa concerns for international researchers. Many Chinese participants have chosen to attend in Mexico City, where interest in DeepSeek’s breakthrough remains high.
FAQs
Q1. What is DeepSeek V3.2-Speciale?
A. It is DeepSeek’s most powerful AI model, designed to match the performance of top global models in reasoning tasks.
Q2. How does V3.2-Speciale compare to Gemini 3 Pro?
A. DeepSeek claims the model equals Gemini 3 Pro in reasoning, although it falls behind in token efficiency and world knowledge.
Q3. Is the model open source?
A. The base V3.2 is open source, while V3.2-Speciale is available through an API.
Q4. Why is the model’s cost lower than Google’s?
A. DeepSeek offers a much cheaper token price to attract developers and increase adoption.
Q5. What challenge does DeepSeek face?
A. The company has limited access to advanced chips, which results in lower compute availability during training.
Also Read- Simple Charging Habits That Improve Your Phone Battery Health
