A Chinese start‑up, DeepSeek, has unveiled its latest model — DeepSeek‑V3.2‑Speciale — claiming that it matches Google DeepMind’s newest Gemini 3 Pro in reasoning performance. Despite limited access to advanced chips, DeepSeek says the new release “equals” Gemini 3 Pro in many reasoning tasks. At the same time, its base model version, DeepSeek‑V3.2, is said to perform on par with GPT‑5 from OpenAI.
According to DeepSeek, V3.2‑Speciale achieved “gold‑medal” level performance on the International Mathematical Olympiad test — a benchmark previously attained only by internal, non‑public AI models from Google or OpenAI. Notably, while the V3.2 model has been open‑sourced on platforms like Hugging Face, the Speciale variant is only accessible via an API, likely due to its high token‑usage and resource demands.
The announcement comes as the global AI community — and major industry events like NeurIPS — witness intensified competition. Many in the AI world have been stirred by DeepSeek’s claim because it suggests a much smaller, less resource‑intensive company can build a model competitive with the best from large, resource‑rich firms.
Still, that doesn’t necessarily settle the debate. Some experts and competing firms point out gaps: while DeepSeek’s engineering and efficiency look impressive, challenges remain in token‑efficiency, general-world knowledge, and robustness compared to top U.S. models. The release — however — marks a significant moment in the global AI race and raises questions about whether the historical dominance of large Western firms remains unassailable.
