DeepSeek released two 685B-parameter models (V3.2 and V3.2-Speciale) on December 1, 2025, claiming frontier performance parity with GPT-5 and Gemini-3.0-Pro on mathematical and coding benchmarks, including gold-medal performance on 2025 IMO and IOI. The models employ a Sparse Attention architecture reducing inference costs 70% on long-context tasks. Both models are released under MIT open-source license with full weights on Hugging Face, eliminating proprietary licensing barriers for deployment. For financial institutions: frontier-capable model access is now available at zero marginal distribution cost globally, restructuring the economics of frontier-model provisioning and raising questions about sustainable pricing and capital-recovery models for proprietary vendors.
420 other tracked developments touching the same claims.