Rumors Emerge About the DeepSeek R2 AI Model Promising 97% Cost Reduction Compared to GPT-4, Exclusively Trained on Huawei’s Ascend Chips
It appears that the Chinese company DeepSeek is preparing to unveil its latest AI model, the “DeepSeek R2.” Recent information has emerged online indicating that this upcoming model could make waves in the market.
DeepSeek R2: A New Player in the AI Scene
DeepSeek made headlines with its initial mainstream model, the R1, which demonstrated that China is making significant strides in AI technology. The launch of R1 was so impactful that it caused a considerable drop in the US stock market, with valuations losing billions. Additionally, DeepSeek’s development approach revealed that creating advanced AI models can be more cost-effective than previously suggested by Western companies such as OpenAI. Now, as per reports from Chinese media, anticipation builds over the forthcoming R2 model, which could provide another surprise for the Western AI market.
🚨Viral rumors of DeepSeek R2 leaked!
—1.2 trillion parameters, 78 billion active, hybrid MoE
—97.3% cheaper than GPT-4 ($0.07/M in, $0.27/M out)
—5.2PB training data. Achieved 89.7% on C-Eval2.0
—Improved vision capabilities at 92.4% on COCO
—82% utilization of Huawei Ascend 910BBig shift away from the US supply chain. pic.twitter.com/Jncg0PvEYU
— Deedy (@deedydas) April 26, 2025
Details on the DeepSeek R2 Model
While the excitement is palpable, it is essential to remember that these reports are based on speculation, and official information from DeepSeek is pending. Chinese sources indicate that the R2 model aims to use a hybrid Mixture of Experts (MoE) architecture, which likely includes more sophisticated gating mechanisms alongside a mix of MoE and dense layers to handle demanding tasks more effectively. The anticipated R2 model is expected to feature approximately 1.2 trillion parameters, which is double that of its predecessor, the R1.
Cost-Effectiveness and Performance
According to these reports, the R2 model could challenge prominent models such as GPT-4 Turbo and Google’s Gemini 2.0 Pro in terms of capabilities and pricing. Specifically, it is said that the cost per token for R2 will be significantly lower than GPT-4, with estimates indicating $0.07 per input token and $0.27 per output token. Compared to OpenAI’s pricing, this pricing strategy positions DeepSeek’s R2 model as a highly attractive option for businesses, potentially revolutionizing the economics of AI technology.
Technical Specifications and Hardware Utilization
Another striking detail regarding the R2 model is its utilization of Huawei’s Ascend 910B chips, achieving 82% efficiency. The computing power is reported to reach 512 PetaFLOPS at FP16 precision. This strategic decision to rely on in-house resources indicates DeepSeek’s commitment to integrating its AI supply chain efficiently.
While it is crucial to take these reports with caution, as final specifications and capabilities of DeepSeek R2 remain unconfirmed, the information from Chinese media suggests that R2 might bring forth substantial changes in the AI landscape, potentially challenging established Western companies.
