DeepSeek-V3-0324 release — MIT license, improved coding/math capabilities, stronger benchmarks

AI Impact Summary

DeepSeek releases an updated base model, DeepSeek-V3-0324, built on the same V3 architecture but now under an MIT license. The update emphasizes improved instruction following, coding and math capabilities, and shows benchmarking gains comparable to GPT-4.5 and above Claude-Sonnet-3.7 on key tasks, signaling a viable open-model alternative for production workloads. Adoption can be accelerated via Hugging Face Inference Providers and TGI/SGLang deployment options, but teams should validate licensing, model card details, and the stability of function-calling and multi-turn rewriting in their prompts before migration.