DeepSeek is reportedly preparing to launch its next-generation AI model, DeepSeek R2, a release that could once again draw global attention to the Chinese AI startup. A few months ago, the company made headlines with DeepSeek R1, a reasoning model that rivaled ChatGPT’s capabilities despite being built with far fewer resources. By optimizing software and using less powerful NVIDIA GPUs, DeepSeek was able to deliver high performance without the extensive hardware typically associated with frontier AI development.
The open-sourcing of DeepSeek R1 further distinguished the company’s approach. Users could install the models locally without relying on cloud services, a move that lowered barriers to access but also raised significant concerns around data privacy and security. Particularly controversial was the discovery that data from DeepSeek’s mobile apps was transmitted to China, prompting advisories to favor local installations over app use. Allegations also emerged accusing DeepSeek of improperly using ChatGPT training data, further complicating its debut.
Now, with DeepSeek R2 reportedly close to release, speculation is mounting about whether it could trigger another market disruption. When R1 debuted, it temporarily shook investor confidence in US tech stocks, exposing vulnerabilities in assumptions about hardware exclusivity as a protective moat for American AI firms. However, broader market forces, including political tensions and tariffs, have since had a much larger impact on financial stability.
According to emerging reports, DeepSeek R2 is expected to offer substantial technical improvements. Built on a hybrid Mixture-of-Experts (MoE) architecture, R2 is said to feature a staggering 1.2 trillion parameters, though only 78 billion are active per token during inference. This design should greatly reduce both training and operational costs. Sources suggest that training DeepSeek R2 was 97.3% cheaper than training GPT-4, with inference costs also dropping significantly—around $0.07 per million input tokens and $0.27 per million output tokens.
Rumors indicate that DeepSeek has pivoted away from relying on NVIDIA, reportedly training R2 using Huawei’s Ascend 910B chips. The company is said to have developed a localized supply chain for its hardware needs, a strategic move that could shield it from supply chain disruptions and international restrictions.
Benchmark results for DeepSeek R2 are rumored to be strong, with particular strengths in reasoning across domains such as finance, law, and patents. Multimodal support is also expected, including sophisticated visual capabilities, further pushing the model toward real-world application across various industries.
While no official release date has been confirmed, signs point to an announcement coming as early as May 2025. If the rumors hold true, DeepSeek R2 could again reshape perceptions of competition in the global AI landscape—though it’s unlikely to cause the same shockwaves in the market that accompanied its predecessor.
