1/ 🐳DeepSeek just dropped Prover-V2-671B, a 671B-parameter AI model built to tackle formal math proofs in Lean 4. It’s open-source, massive, and optimized for theorem proving—think of it as the mathlete of LLMs. Let's dive in 🧵👇

🤖 Software Dev by day and AI & Crypto
⭐ I love sharing all the new tech developments with you
📫 DM me for collaboration
🌐 AI MARKET WATCH --> t.co/WMNyhRhSAO

2/ What's different about it?  It was trained using a recursive pipeline and employs a Mixture-of-Experts (MoE) architecture:  Prover-V2 learns to sew problems into complete proofs when DeepSeek-V3 breaks them down into subgoals.

3/ Benchmarks? It hits an 88.9% pass rate on MiniF2F and solves 49 out of 658 PutnamBench problems. It even tackled 6 out of 15 AIME competition problems. That’s serious math cred. It’s not just symbolic manipulation—it’s doing real, verifiable reasoning.

5/ DeepSeek-Prover-V2-671B is a game-changer for formal math. It’s open, powerful, and pushing the boundaries of what AI can do in theorem proving. If you’re into math, logic, or just cool AI stuff, this is worth checking out.

https://huggingface.co/deepseek-ai/DeepSeek-Prover-V2-671B

https://arxiv.org/abs/2504.21801​

https://github.com/deepseek-ai/DeepSeek-Prover-V2

4/ You can run it via Hugging Face or OpenRouter. It supports up to 163K tokens of context, so it can handle long, complex proofs. There’s also a 7B version if you don’t have a GPU farm lying around. Both are free to use.