DeepSeek rumored to build R2 AI model using Huawei AI chips

Looks like DeepSeek is now developing another AI model – R2, powered by Huawei Ascend 910B chips. The company has recently taken center stage for its cost-effective LLM and now once again caught attention for the next-gen logical AI tech.

X blogger @deedydas shared viral rumors of the DeepSeek R2 AI model on his page. The post reveals the key specifications and features of the new AI technology.

As per the details, DeepSeek R2 will use a hybrid MoE (Mixture of Experts) architecture, one of the advanced versions of existing MoEs, and can offer an advanced gating mechanism + dense layers to improve high-end AAI workloads.

DeepSeek R2 may further double the parameters over R1, with around 1.2 trillion parameters. Interestingly, it can be 97.3% cheaper than ChatGPT-4 for enterprise use at $0.07/M input token and 0.27/M output tokens.

If it’s true, the DeepSeek R2 AI model will be the most cost-effective LLM on the market, giving a bang to its foreign opponents like GPT-4 Turbo and Gemini 2.0.

Although the highlight is that the rumored DeepSeek R2 AI model is relying on Huawei Ascend 910B chips. Inputs show that the model has achieved 82% utilization of the Ascend 910B processor cluster, and almost fully trained on Huawei chips.

910B also takes the computing power of the R2 model to 512 PetaFLOPS of FP16 precision. This shows that DeepSeek is depending on home-grown resources for its new and powerful AI models. You can check the key features below:

• 1.2 trillion parameters, 78B active, hybrid MoE
• 97.3% cheaper than GPT 4 ($0.07/M in, $0.27/M out)
• 5.2PB training data. 89.7% on C-Eval 2.0
• Better vision. 92.4% on COCO
• 82% utilization in Huawei Ascend 910B

DeepSeek rumored to build R2 AI model using Huawei AI chips (Image Credits: Deedydas/X)

It’s not the first time that DeepSeek (a Chinese AI startup) has used Huawei AI chips for its model. The company has launched the first Ascend 910B-powered R1 model in January this year, challenging OpenAI and Google, with three key factors:

• Cost-effectiveness
• Open-Source
• Efficiency

Now it is seemingly bringing a new model to the market with probably even better features and functions to simplify the AI use for customers, making it easier.

Do note that these details are still rumors, and there is no official confirmation on this matter. Hence, we suggest taking this input with a grain of salt at the moment.

|| source ||