金工周报-20250729

• NVIDIA launched the OpenReasoning-Nemotron reasoning model series in July 2025, based on the Qwen2.5 architecture, distilled from the 671 billion-parameter DeepSeek R1 0528 model, and available in four parameter scales: 1.5B, 7B, 14B, and 32B. The model aims to support structured tasks such as mathematics, science, and code generation efficiently [12] - The core innovation of OpenReasoning-Nemotron lies in its data distillation strategy, leveraging the NeMo Skills framework to generate 5 million high-quality data trajectories covering mathematical proofs, scientific derivations, and programming solutions. The training process uses supervised fine-tuning (SFT) instead of reinforcement learning, ensuring logical consistency and precision in symbolic reasoning [12] - The model employs the GenSelect algorithm to implement a "heavy reasoning mode," which involves parallel generation of candidate solutions by multiple agents and selecting the optimal answer. For example, the GenSelect@64 on the 32B model improved HMMT math competition scores from 73.8 to 96.7 and enhanced LiveCodeBench scores from 70.2 to 75.3 in code generation tasks [13] - The OpenReasoning-Nemotron series achieved record-breaking results in benchmarks such as GPQA, MMLU-PRO, and AIME24. The 32B model scored 89.2 on AIME24, surpassing OpenAI's o3-high model, while the 7B model scored 78.2, representing a nearly 20% improvement over its predecessor. However, the 1.5B model showed performance degradation to 45.6 due to inconsistencies in handling 32K tokens [15] - The Qwen3-Coder model, developed by Alibaba Cloud's Tongyi Qianwen team, was officially open-sourced in July 2025. It features a 480 billion parameter scale with a native 256K context window and employs a sparse MoE design, activating only 35 billion parameters per inference. The model was trained on a 7.5 trillion token corpus, with 70% of the data being code, covering over 80 programming languages and 20 markup languages [19][20] - Qwen3-Coder achieved a HumanEval pass@1 accuracy of 93.7%, surpassing Claude 3.5's 92.4%. On the SWE-Bench Verified benchmark, it achieved a 31.4% task success rate, exceeding GPT-4's 30.9%. Key innovations include extending the native 256K context to 1M tokens using YaRN technology and integrating execution feedback mechanisms to validate and reward generated code [20] - The GitLab Duo platform, launched in public beta in July 2025, virtualizes traditional software development team roles into specialized AI agent clusters. These agents handle tasks such as requirement planning, code writing, security analysis, testing, and operations management, forming a dynamic collaboration network. The platform automates workflows through the "Flows" feature, enabling developers to input functional descriptions and have agents complete tasks like requirement decomposition, code generation, and testing [33][36] - GitLab Duo integrates with mainstream development environments like VS Code and JetBrains IDEs and plans to introduce a "knowledge graph" feature to enhance agents' understanding of code context. The platform also emphasizes security, employing end-to-end encryption and sandbox environments for code validation [36][37]