Data Machina #258

Substack · 2 years ago · 11

DeepSeek-Coder-v2 became the top open-source code model matching GPT-4 Turbo performance, while multiple other advances emerged including Hermes 2 Theta Llama-3 70B surpassing Llama-3 on benchmarks, Unique3D generating 3D meshes from single images, and Claude 3.5 Sonnet release with strong vision and coding capabilities. Key benchmarks include Hermes 2 Theta matching GPT-4 on MT Bench and DeepSeek-Coder-v2 achieving GPT-4 Turbo-level code performance. These releases lower barriers to accessing capable models and introduce new capabilities in video generation, vision tasks, and instruction following, with some models made freely available or open-sourced.