新京报贝壳财经讯（记者罗亦丹）2月25日，继除夕开源Qwen3.5-397B-A17B之后，阿里继续开源千问 3.5系列模型。本次开源三款中等规模的新模型，包括Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、 Qwen3.5-27B。基于架构创新和训练突破，此次开源的三款千问3.5模型均创下中等尺寸模型的性能新 高，超越了更大尺寸的上代旗舰模型Qwen3-235B-A22B和Qwen3-VL，多榜单表现均明显优于GPT-5 mini。 编辑 岳彩周 校对 柳宝庆 值得注意的是，千问3.5新模型甚至可直接部署于消费级显卡，对开发者极为友好。目前，基于 Qwen3.5-35B-A3B的托管模型Qwen3.5-Flash已上线阿里云百炼，每百万Token输入低至0.2元。千问3.5模 型采用混合注意力机制，结合高稀疏的MoE架构创新，并基于更大规模的文本和视觉混合Token上训 练，新模型以更小的总参数和激活参数量，实现性能提升。 ...

Core Insights - Alibaba's Qwen announced the open-source release of a high-efficiency mixture of experts (MoE) model, Qwen3-Coder-Next, designed for programming agents and local development, featuring a total of 80 billion parameters with only 3 billion activated per inference [1] Group 1 - The Qwen3-Coder-Next model is available in two versions: Qwen3-Coder-Next (Base) and Qwen3-Coder-Next (Instruct), supporting research, evaluation, and commercial applications [1]

Core Viewpoint - The company Zhipu has officially released and open-sourced GLM-OCR, a model with a parameter size of only 0.9 billion, achieving state-of-the-art (SOTA) performance in various mainstream benchmarks for formula recognition, table recognition, and information extraction [1] Group 1: Model Features - GLM-OCR is optimized for scenarios including handwriting, complex tables, code documents, seal recognition, and multilingual mixed typesetting [1] - The model employs an "encoder-decoder" architecture, integrating a self-developed CogViT visual encoder, and utilizes a two-stage technical process of "layout analysis → parallel recognition" [1] Group 2: Performance and Efficiency - The model can process PDF documents at a throughput of 1.86 pages per second [1] - The pricing for API calls is set at 0.2 yuan per million tokens [1] Group 3: Deployment and Tools - GLM-OCR supports deployment on vLLM, SGLang, and Ollama [1] - The complete SDK and inference toolchain for the model have been open-sourced, making it suitable for high concurrency and edge computing scenarios [1]

Core Insights - The article discusses the recent launch of GLM-4.7 by the company, which has submitted its prospectus to the Hong Kong Stock Exchange [1] - GLM-4.7 enhances coding capabilities, long-range task planning, and tool collaboration, achieving leading performance in several mainstream public benchmark tests [1] Group 1: Model Enhancements - GLM-4.7 has made breakthroughs in programming, reasoning, and agent dimensions, significantly improving its performance in multilingual coding and terminal agents [1] - The model now implements a "think first, act later" mechanism in programming frameworks such as Claude Code and TRAE, demonstrating more stable performance on complex tasks [1] Group 2: API and Development Integration - GLM-4.7 is available through an API provided by BigModel.cn and has been integrated into the z.ai full-stack development mode with a new Skills module [1] - The Skills module supports unified planning and collaborative execution of multimodal tasks [1]

Core Insights - Tongyi's first deep research agent model, DeepResearch, has been officially open-sourced, featuring a parameter size of only 30 billion (with 3 billion activated), achieving state-of-the-art (SOTA) results across multiple authoritative evaluation sets, surpassing many top agent models [1][5] Model Training - The Tongyi team has developed a complete training pipeline driven by synthetic data, integrating pre-training and post-training phases. This model capability is based on a multi-stage data strategy aimed at creating vast amounts of high-quality training data without relying on expensive manual annotations [3] - The training pipeline is optimized based on the Qwen3-30B-A3B model, incorporating innovative reinforcement learning (RL) algorithms for validation and real training, enhancing model efficiency and robustness. The use of asynchronous reinforcement learning algorithms and automated data curation processes significantly boosts the model's iteration speed and generalization ability [3] Model Performance - The DeepResearch model, with 3 billion activated parameters, performs comparably to flagship models such as OpenAI's o3, DeepSeek V3.1, and Claude-4-Sonnet in various authoritative agent evaluation sets, including Humanity's Last Exam (HLE), BrowseComp, and GAIA [5] Model Applications - The model has been applied in various real-world scenarios, such as the "Xiao Gao Teacher" developed in collaboration with Amap, which acts as an AI co-pilot for complex travel planning tasks. Additionally, Tongyi's legal research agent, empowered by the DeepResearch architecture, can autonomously execute complex multi-step research tasks, simulating the workflow of a junior lawyer [7] DeepResearch Agent Series - Tongyi DeepResearch also boasts a rich family of DeepResearch Agent models. Earlier this year, the team has continuously expanded its DeepResearch offerings, with previously open-sourced models like WebWalker, WebDancer, and WebSailor achieving industry-leading results in agent synthetic data and reinforcement learning [9]

Core Insights - The first deep research Agent model, DeepResearch, developed by Tongyi, has been officially open-sourced [1] - The model has 30 billion parameters (with 3 billion activated) and has achieved state-of-the-art (SOTA) results on multiple authoritative evaluation datasets [1] Company Summary - Tongyi has introduced its first deep research model, indicating a significant advancement in AI technology [1] - The open-sourcing of DeepResearch may enhance collaboration and innovation within the AI research community [1] Industry Summary - The achievement of SOTA results by DeepResearch highlights the competitive landscape in AI model development [1] - The model's parameters and performance metrics suggest a trend towards larger and more efficient AI models in the industry [1]

Core Insights - Tencent Hunyuan has launched four small-sized models: 0.5B, 1.8B, 4B, and 7B, all of which are open-source and can run on consumer-grade graphics cards [1] - The models are compatible with major chip platforms including ARM, Qualcomm, MTK, and Intel, making them suitable for various devices such as smartphones, computers, smart cars, and smart home devices [1] - The new models have demonstrated leading performance in language understanding, mathematics, and reasoning during testing [1] Model Specifications - Hunyuan-4B: - Low power consumption and high efficiency - Maximum input and output of 32K - Suitable for real-time response scenarios with strong inference performance and accuracy - Supports various verticals like finance, education, and healthcare with cost-effective fine-tuning [2] - Hunyuan-7B: - High cost-performance ratio - Can run efficiently on consumer-grade GPUs - Knowledge density significantly higher than other models of similar size, approaching the performance of larger models - Offers flexibility in balancing response rate and depth, with a 256K ultra-long context window [2]