DeepSeek’s new math model stirs buzz about its mysterious next-gen LLM, R2

Chinese AI startup DeepSeek has dropped a surprise upgrade to its math-focused language model, intensifying speculation around its upcoming next-generation reasoning system known simply as R2.

While the company has remained tight-lipped about the new model, the sudden release of Prover-V2, a 671-billion-parameter model fine-tuned for mathematical proof-solving, has reignited online chatter across developer and investor communities alike.

The new model, based on DeepSeek’s V3 foundation, was quietly open-sourced on Wednesday (April 30). It builds on Prover-V1.5, which launched last August and drew interest from academia and competitive math circles.

While Prover-V2 is not the long-awaited R2, it has been widely interpreted as a key stepping stone. Users on X and Reddit are calling it a math ability upgrade laying the groundwork for what could be the next leap in reasoning-focused LLMs from China’s most-watched AI startup, South China Morning Post reported.

More from Tech

China's DeepSeek is ready with version 2. What to expect Explained: The controversy over Pakistan military officer's throat-slit gesture at Pahalgam attack protesters in London

Founded in 2023 by Liang Wenfeng as a spinout of his quantitative hedge fund High-Flyer, DeepSeek quickly gained global attention with its R1 model, launched in January. R1 stunned the AI world by matching OpenAI’s o1-level performance at a fraction of the cost, all while using far fewer resources. That success set expectations sky-high for whatever comes next.

No timeline for R2

However, DeepSeek has offered no public timeline for R2. The company has revealed little beyond research papers and model updates, fuelling a vacuum of information that has been filled by social media speculation. One viral post from a DeepSeek researcher simply announcing Prover-V2 led to a cascade of replies pleading for an R2 release. “R2 R2 R2 please,” one user wrote.

Even more buzz came from Chinese stock-trading forums like Jiuyangongshe, where rumors of an imminent R2 drop spilled over into Western platforms. A notable US venture capital investor picked up the chatter on X, propelling the news into wider investor circles. Searches for “DeepSeek” and “R2” have spiked on Google Trends over the past week.

Impact Shorts

More Shorts

America ready for self-driving cars, but it has a legal problem

Alibaba, Baidu begin using own AI chips as China shifts away from US tech amid Nvidia row

Adding to the intrigue, DeepSeek is now quietly ramping up hiring. The company recently posted openings for its first product and design lead, based in either Beijing or Hangzhou. The job description calls for building a “next-generation intelligent product experience” rooted in LLM tech. The startup is also actively recruiting a chief financial officer and chief operating officer.

Competition in China rising

This comes just as other major Chinese firms are upping their game. On Tuesday, Alibaba unveiled Qwen3, its latest family of models that the company says outperform DeepSeek-R1 on several metrics. The announcement was seen by some as a shot across the bow, upping the pressure on DeepSeek to deliver a follow-up.

Meanwhile, in the United States, OpenAI recently released o3 and o4-mini, touting them as its “most capable models to date.” While DeepSeek lacks access to cutting-edge Nvidia chips due to US export restrictions, it has built a reputation for maximising performance on constrained hardware, drawing interest from technologists and policymakers alike.

The launch of Prover-V2 may not be the generational leap that some were hoping for, but it suggests DeepSeek is far from idle. With the company scaling up and hype building fast, the question now is not whether R2 is coming, but how close we are to seeing it in action.