Articles

DeepSeek officially released two landmark Large Language Models: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. DeepSeek-V3.2 aims to balance strong reasoning capabilities with efficient output length, making it particularly suitable for daily Q&A and general Agent tasks. It has reached GPT-5 level in public reasoning benchmarks, while significantly reducing computational costs and user waiting time compared to similar models. DeepSeek-V3.2-Speciale maximizes the reasoning capabilities of open-source models. By combining the theorem proving capabilities of DeepSeek-Math-V2, it performs comparably to Gemini-3.0-Pro on mathematics, code, and general domain evaluation sets, and has won multiple international Olympiad gold medals. The article emphasizes an innovation of DeepSeek-V3.2: the first deep integration of thinking mode with tool calling. Through a large-scale Agent training data synthesis method, the model's generalization ability on complex tasks has been greatly improved, enabling it to reach the highest level of open-source models in Agent evaluations. Both models are open-source and provided services through the official web interface, App, and API. V3.2-Speciale also provides temporary API services for community evaluation and research.

The article discusses in detail the origin and development of Large Language Model (LLM)'s 'Deep Thinking' ability, starting from the essence of machine thinking, the microstructure of the Transformer architecture, and the importance of parameter scale. Next, it explains the necessity of deep thinking in mitigating hallucinations and introduces external guidance methods such as CoT and its variants in detail. The core part deeply analyzes Reinforcement Learning from Human Feedback (RLHF), compares the advantages and disadvantages of mainstream algorithms such as PPO, DPO, and Group Relative Policy Optimization (GRPO), and provides a critical review of the limitations of RLHF. Finally, it looks forward to future cutting-edge directions such as Atomic Thinking, Hierarchical/Recursive Reasoning Model (HRM/TRM), and Training-Free GRPO, providing profound insights and multiple possibilities for the future development of LLMs.

This article addresses the critical bottleneck of context management in scaling sophisticated AI agents for long-horizon tasks, arguing against merely increasing context window size. It proposes "context engineering" as a new discipline, treating context as a first-class system with its own architecture and lifecycle. Google's Agent Development Kit (ADK) is introduced as an open-source, multi-agent-native framework built on the thesis that context is a "compiled view" over a richer stateful system. ADK's architecture adheres to three core principles: separating storage from presentation, explicit transformations via named processors, and default context scoping. The framework details a tiered context model comprising working context, session (a durable log of structured events), memory (long-lived, searchable knowledge), and artifacts (large external data). It emphasizes pipeline-based context processing, compaction, caching, and agent-directed relevance management for artifacts and memory. For multi-agent systems, ADK provides explicit scoping mechanisms and conversation translation to prevent context explosion and maintain agent coherence during handoffs, moving beyond ad-hoc string manipulation to a rigorous systems engineering approach for production AI agents.

This article analyzes the working principles of RAG (Retrieval-Augmented Generation) systems, explaining the entire process from document chunking and vectorization to semantic retrieval and answer generation through clear analogies. Then, using 'The Lychee of Chang'an' as an example, the article demonstrates how to use the LlamaIndex data framework to quickly build an intelligent question answering system with concise Python code and provides a detailed interpretation of the core APIs. The article also explores the impact of key parameters such as chunk_size, top_k, and chunk_overlap on question answering quality through rigorous experiments. It provides parameter tuning suggestions for various application scenarios. Finally, the article dissects the internal architecture of LlamaIndex to help readers understand its operation and prepare for further customization. The overall content combines theory and practice, making it highly instructive.

DeepSeek has recently released two official models: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. DeepSeek-V3.2 aims to balance reasoning ability and output length, making it suitable for daily Q&A and general Agent-based tasks. It performs excellently in public reasoning benchmarks, approaching GPT-5 level. DeepSeek-V3.2-Speciale pushes the reasoning ability of open-source models to the extreme, integrating the theorem-proving capabilities of DeepSeek-Math-V2, rivaling Gemini-3.0-Pro in complex tasks such as mathematics and code, and successfully winning multiple international competition gold medals. A major innovation highlighted is that DeepSeek-V3.2 is the first model to integrate reasoning into tool usage. Through large-scale Agent training data synthesis, the model's generalization ability in agent evaluation is significantly improved, narrowing the gap between open-source and closed-source models. The release includes open-source links for both models, updates to the official web page, App, and API, and a guide to using the API for tool usage in reasoning mode.

AI large language model technology is becoming increasingly popular. However, enterprises face two major challenges when implementing AI applications: the lack of private data and the lack of memory in models. To address these challenges, the article points out that databases need to evolve to store enterprise private data and user behavior in real-time, and support efficient context engineering (the process of preparing data for AI models). The author proposes a new database workload - 'Agent-Oriented Multi-Modal Hybrid Search', which requires the database to handle structured, semi-structured, and unstructured data, and support various indexes such as relational, full-text, vector, and graph. The article elaborates on the evolution and core principles of AI databases, emphasizing the integration of multi-modal hybrid search while retaining traditional HTAP capabilities. In terms of implementation path, the author prefers extending AI capabilities on top of mature relational databases. The article delves into the necessity of AI-native hybrid search, pointing out that single vector search is not sufficient to meet the needs of complex AI applications, and it is necessary to combine multiple search methods and re-rank them. At the same time, modern data architectures should be easy to use, flexible (Serverless, distributed), and future-oriented (natively support AI workloads, massive multi-tenancy). Finally, the article envisions the 'AI Native' database form with deep integration of data and models, realizing the direct processing of unstructured data and model invocation by the database kernel through technologies such as AI Function. Based on these considerations, OceanBase has open-sourced seekdb, an AI-native database for developers, aiming to provide a lightweight, all-in-one AI database solution to simplify AI application development.

The article deeply analyzes the four generations of LLM 'thinking' ability: from 'no thinking' to 'thinking', then to 'tool calling Agents', and finally to 'interleaved thinking Agents'. It explains that the essence of interleaved thinking is a dynamic real-time feedback loop between thought and action, enabling models to 'think, act, and correct' simultaneously, emulating human-like behavior. The article also highlights the key role of DeepSeek-V3.2 and MiniMax M2 in advancing this technology and its ecosystem adoption (including open-source frameworks and toolchain adaptation), showcasing significant performance gains with examples and data.

This article provides a detailed analysis of the technical report "DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models." The article points out that DeepSeek-V3.2 has reached parity with GPT-5-High in reasoning capabilities, and its high-computing power version, Speciale, is close to Gemini-3.0-Pro in mathematics and programming. The key to achieving these breakthroughs lies in three core technologies: First, **DeepSeek Sparse Attention (DSA)** reduces the computational complexity of long contexts from O(L²) to O(Lk), significantly improving **long-context processing capabilities** and efficiency without sacrificing performance; second, **Large-Scale Post-Training**, with a computing budget exceeding 10% of that of pre-training, effectively avoids catastrophic forgetting in multi-stage training through expert distillation and GRPO hybrid RL training, and **Significantly Improved Model Performance on Challenging Tasks**; finally, **Large-Scale Agent Data Synthesis**, by generating 1,800 environments and 85,000 "difficult-to-solve, easy-to-verify" tasks, greatly improves the model's generalization ability in agent scenarios. The article also elaborates on the optimization of V3.2 in **Thinking in Tool-Use**, meaning that reasoning content is not discarded when the tool returns results. This improves the efficiency of tool invocation. Finally, the report candidly points out the limitations of DeepSeek-V3.2 in world knowledge, **Token Efficiency**, and the most difficult tasks.

The article is a compilation of an interview with Martin Fowler on The Pragmatic Engineer program. He begins by reviewing his personal career and sharing the operating mechanism of ThoughtWorks' Technology Radar. The core content focuses on how AI is changing software development. It draws a parallel to the transition from assembly language to high-level languages but emphasizes the paradigm shift from deterministic to non-deterministic programming. He points out the great potential of AI in rapid prototype development and understanding legacy systems, but also warns of the drawbacks of AI-assisted rapid prototyping (where code quality is difficult to guarantee) and the unreliability of AI output, emphasizing the importance of rigorous testing and refactoring. Fowler also discusses the new challenges of agile development in the AI era and expresses concerns about the current tech industry layoff wave and AI bubble. Finally, he provides career development advice for junior engineers, emphasizing the core values of curiosity, in-depth study, and excellent communication skills.

The article details the distributed optimization practice of Tencent's 'Yinian LLM' inference framework. Faced with the significant gap between the performance of current LLM inference frameworks and their theoretical limits, as well as business demands for rapid, stable, and efficient systems, the 'Yinian' framework was developed. Its core design philosophy involves hand-written C++ models and full-process autonomous memory management (especially KV Cache). This enables unified support for heterogeneous hardware such as Nvidia, Huawei Ascend, and Tencent Zixiao. The article deeply analyzes the performance bottlenecks in the Prefill and Decode stages and proposes various solutions such as MoE load balancing, MLA optimization, and multi-dimensional parallelism (TP/DP/EP). It particularly emphasizes the key role of PD Separation (Prefill and Decode Separation) and multi-stage pipeline parallelism (PP) in reducing communication overhead, improving memory utilization, and throughput. Finally, it looks forward to future optimization directions in scheduling strategies, flexible KV Cache, layer decoupling, and cross-Batch process orchestration, providing a comprehensive engineering perspective and practical experience for LLM inference systems.