XNews

New Models and Benchmarks: DeepSeekMath-V2 and Whisper Thunder

New Models and Benchmarks: DeepSeekMath-V2 and Whisper Thunder

Core Events: DeepSeek releases DeepSeekMath-V2 and makes it open-source; an anonymous video model, Whisper Thunder, ranks first on the leaderboard. DeepSeekMath-V2 performs excellently on benchmarks such as IMO-ProofBench, IMO 2025 (some problems), and Putnam 2024, achieving a Putnam score close to 118/120, reaching the level of a gold medal at IMO 2025. DeepSeekMath-V2 is open-sourced under the Apache 2.0 license and available for download on GitHub and Hugging Face. A mysterious video model named Whisper Thunder (also known as David) ranks first on the Artificial Analysis leaderboard, surpassing known models such as Veo 3/3.1, Sora 2 Pro, and Kling 2.5, with its ownership currently unknown.

Key insights

DeepSeekMath-V2 Approaches Human Gold Medal Level in Math Competition Benchmarks

DeepSeek's released DeepSeekMath-V2 shows outstanding performance in the IMO-ProofBench, actual competitions (such as some problems from IMO 2025), and Putnam 2024, achieving a Putnam score close to 118/120, described as reaching the gold medal level of IMO 2025. This model is open-sourced under the Apache 2.0 license and available for download on GitHub and Hugging Face.

Whisper Thunder Becomes the Top-Ranked Mysterious Video Model

An anonymous video model known as Whisper Thunder (also referred to as David) ranks first on the Artificial Analysis leaderboard, surpassing known models such as Veo 3, Veo 3.1, Sora 2 Pro, and Kling 2.5. The ownership and origin of this model have not yet been disclosed.

Source posts