DeepSeek
5 articles found in this topic.
DeepSeek V3.2's Agentic Performance Surges 40% with Interleaved Thinking
DeepSeek V3.2 significantly boosts its agentic capabilities by 40% using Interleaved Thinking, a method that combats "state drift" in large language models. This approach involves alternating between reasoning and tool calling, explicitly recording thought processes to maintain long-term plans and improve performance across various tasks.
OpenRouter Study Maps 100 Trillion Tokens of AI Usage in 2025
OpenRouter and a16z study 100T AI tokens to reveal how models are used in 2025. Open-source and Chinese models surge as medium-sized systems become core tools.
Open-Source Models Advance OCR Workflows
The emergence of Vision-Language Models (VLMs) is significantly expanding OCR capabilities beyond simple text extraction to complex visual and semantic understanding. Open-source models offer cost-efficiency and privacy benefits, making advanced OCR solutions more accessible. This article explores key factors for selecting OCR models and highlights cutting-edge open-source options.
DeepSeek Releases V3.2 Models with Enhanced Agent Capabilities and Integrated Reasoning
DeepSeek has launched DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, enhancing agent capabilities and integrating advanced reasoning. The V3.2 model balances reasoning with output length for daily tasks, while Speciale pushes open-source boundaries with theorem proving and achieves top-tier benchmark results.
DeepSeek Introduces Advanced Models, Challenging Proprietary AI Performance
DeepSeek has released DeepSeek V3.2 and DeepSeek-V3.2-Speciale, aiming to close the performance gap with proprietary AI. These new models, featuring Sparse Attention and enhanced post-training, compete with GPT-5 and Gemini, securing top spots in competitions.