5 DeepSeek R2 Rumors (A Summary Guide)

People are expecting new advancements of AI models. Beyond doubt, DeepSeek is at the top of the list.

Since its R1-0528 version released on March 28th, the reasoning depth shows significant changes. And AI enthusiasts excitement grew. They started guessing wildly about the new version - DeepSeek R2.

Rumors rise.

1. The Release Time

In March 2025, online rumors suggested an early March 17th release for DeepSeeke R2. However, a claim swiftly denied by the official team.

Then in April, still, whispers persist online about DeepSeek R2 will potentially arrive as early as at the end of April. As it stands, this news is untrue.

Now it's June. DeepSeeker R2 hasn't appeared yet.

So far, there has been no official announcement of the release date of DeepSeek R2.

2. Architecture Changes

Unconfirmed online reports indicate that DeepSeek R2 may depart from the traditional Transformer architecture. And it will use Recursive Cognition Lattices to instead.

The official has not confirmed the news.

Another rumor is that DeepSeek R2 will use Hybrid Mixture of Experts (MoE) 3.0 architecture. This will reduce reasoning cost as low as 2%-5% of GPT-4.

3.The Chip Used

It's said that DeepSeek R2 is using Huawei AI chip (Ascend 910B), not Nvidia.

The utilization rate of chip reaches 82%, and the performance is close to 91% of A100

If this rumor is true, this signals a major move towards hardware diversification, lessening reliance on Nvidia GPUs.

4.Costs May Drop Sharply

One rumor says, the input token will be reduced to $0.07 per million tokens while $0.27 per million for output tokens.

That is to say, there is a 97.3% decrease compared to GPT-4 Turbo.

So, will we see an explosion of AI applications?

5. Enables Deployment on Consumer-Grade Devices

A rumor reports that DeepSeek R2 can see an 83% reduction in model size achieving by 8-bit quantization compression, and sacrificing less than 2% in accuracy.

This may bring out: reasoning speed improved by 40%, and energy use dropped by 25%, making it ideal for edge applications like smart homes and self-driving cars.