智东西（公众号：zhidxcom）
作者 | 王涵
编辑 | 李水青
智东西10月28日报道，今天，深圳大模型厂商扩散智能（DiffuSpace）发布并开源扩散语言模型Clover-35B-A3B-Preview，该模型采用混合专家架构（MoE），总参数350亿、激活参数30亿，支持思考模式。
该公司发布的基准测试结果显示，Clover-35B-A3B-Preview在SWE-bench Verified、Terminal-Bench2.1等6项智能体编程及工具使用任务评测中，排名同级别模型中的第一，超过同规模自回归模型Qwen3.6-35B-A3B。
HuggingFace：
https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
GitHub：
https://github.com/DiffuSpace/Clover-35B-A3B-Preview
ModelScope：
https://modelscope.cn/models/Diffuspace/Clover-35B-A3B-Preview
申请在线体验：https://chat.diffuspace.ai/
同时，扩散智能还推出了Clover-35B-A3B-Preview的决策模型，面向快速、结构化的判断任务，支持独立模式和联合模式两种作答方式，该模型将在后续发布。
不同于TypeSafe的Jev等开源决策模型的单点评估思路，联合模式以全局统筹的方式开展推理，省去多次交互的耗时，使多项判断相互参照、规避逻辑冲突，一次性完成一组关联决策，实现更快、更稳定的推理结果。
在包含500组、2420道关联问题的内部验证中，联合模式答对95.0%，并在80.4%的问题组中全部答对。相比之下，同一模型独立作答时，两项成绩分别为89.8%和62.4%；Jev在相同题目的历史评测中分别为63.6%和15.8%。
在对比案例中，针对合同中的判断问题，Clover-35B-A3B-Preview决策模型的联合模式的5项判断全部正确，独立模式答对4项，Jev答对2项。
该决策模型还和Jev展开了一场中国象棋的对弈，结果是扩散智能的决策模型执黑获胜。
扩散智能于2026年成立于深圳，由港大NLP实验室联合主管孔令鹏教授，以及其博士生龚珊三、叶佳成联合创立。核心团队长期研究扩散语言模型，是全球最早探索该路线的团队之一，在条件生成、离散信号建模、Scaling三大dLLM（扩散大语言模型）核心问题上取得全球领先突破。
该团队早在2022年就启动dLLM前沿研究，提出了DiffuSeq，开创了扩散模型处理序列到序列文本生成（Seq2Seq）的技术路径。2025年4月，其推出扩散语言模型Dream 7B，至今Dream系列模型累计在Hugging Face下载超270万次。
10月9日，扩散智能首次公布融资，其连续完成两轮融资，总金额数亿元人民币，创下公开披露的全球dLLM最大规模融资纪录。
一、6项Agent测评第一，单卡GB300生成速度达680 tokens/s
在基准测试方面，Clover-35B-A3B-Preview在智能体编程与工具使用任务中表现较好。在本次对比的6项Agent评测中，该模型均取得了所列模型中的最高分。
其中，SWE-bench Verified达到73.6分，Terminal-Bench2.1达到51.7分，τ³ Overall达到65.3分，超过了同为扩散语言模型的LLaDA2.2 flash、Mercury 2.5，以及同尺寸的自回归模型Qwen3.6-35B-A3B和Qwen3.5-35B-A3B。
在指令遵循、知识、逻辑推理和长文本理解等多项评测中，其成绩达到或超过同规模自回归模型，GPQA Diamond、ZebraLogic和AA-LCR分别达到83.8、95.8和72.0分。
在推理效率方面，团队使用覆盖代码编写、终端操作和工具调用等场景的内部Agent任务执行记录进行测试。
端侧部署方面，Clover-35B-A3B-Preview在单张NVIDIA GB300 GPU上，以BF16精度、batch size=1，达到680 tokens/s的整体生成速度。
此外，在单张NVIDIA H200 GPU上，Clover达到515tokens/s，高于DFlash2方案的365 tokens/s。
与其他代表性轻量级、高效推理或端侧部署模型的对比中，Clover-35B-A3B-Preview在9项评测中的6项取得最高分，优势覆盖知识推理、工具使用与智能体编程。
二、路线规划28步完成，49项Agent任务耗时中位数仅1.52秒
基于Clover-35B-A3B-Preview在推理、工具使用以及编程上的能力，用户可以借助该模型完成路径规划、WebMCP任务、模拟桌面环境、智能体调度等工作。
例如，用户要查询“从匹兹堡CMU到最近的Wendy’s需要开车多久”。Clover识别出结果偏离目标，逐步加入城市、街区和邮编改写检索，并在页面卡住后两次重置、重填路线，最终找到目标门店，查得驾车距离 2.0公里、车程约3分钟。
本次对照中，Clover-35B-A3B-Preview用28步、169.9秒完成任务；Qwen3.6 Flash用满30步、耗时578.3秒，仍停留在距起点347公里、车程约4.5小时的门店，未通过任务评判。
通过网页提供的工具，Clover-35B-A3B-Preview可以连续完成开票、课程注册、金额汇总、事件状态与客户信息查询等任务。
在完整的49项任务、每项3次尝试的评测中，Clover的Agent执行耗时中位数为1.52秒，Gemini 3.6 Flash为8.86秒。
此外，Clover-35B-A3B-Preview还可以创建私有项目并邀请成员。在相同的智能体框架、提示词、工具和浏览器配置下，本次运行中Clover-35B-A3B-Preview用13步、169.9秒完成任务，Qwen 3.6 Flash用21步、392.3秒完成。
该案例中用户要求用单个HTML文件模拟一个桌面环境，支持窗口拖动，并能打开多个小应用。
Clover-35B-A3B-Preview生成了包含程序坞、窗口系统以及记事本、计算器、时钟、天气和设置等应用的模拟桌面。
在演示中，用户可以打开应用、输入文本、操作计算器并调整设置。
在城市低空无人机调度演示中，模型通常需要将复合空管报文解析为禁飞区解除、航点修改、任务分配、禁飞区半径调整和紧急调度等多项任务。
Clover-35B-A3B-Preview在0.97秒后同时完成处理呈现5项正确结果。Qwen在3.78秒展示结果，还将UAV5的测绘任务误解析为医疗任务。
三、重构双向门控Delta网络，兼顾推理速度与性能
Clover-35B-A3B-Preview能够兼顾推理速度与性能表现，背后是扩散智能人团队针对性的架构优化。
在Clover-35B-A3B-Preview中，扩散智能团队完成了双向门控Delta网络（Bidirectional GDN）的重构设计，针对历史记忆与块内去噪两条链路采用差异化计算范式。
模型已生成的上下文信息依托GDN递归状态实现高效压缩，当前块内Token则借助双向Softmax注意力完成细粒度交互。
在核心机制上，技术团队将历史记忆读出作为独立信息单元，与块内Token纳入同一次Softmax运算开展信息筛选，由模型自主学习在推理过程中对历史信息与当前块信息的依赖权重。
团队同时新增可学习块大小校准机制，用于缓解块内Token数量增加带来的历史权重稀释问题。
针对局部特征提取环节，扩散智能团队还为右侧上下文搭建了独立的零初始化卷积分支。该方案基于预训练因果卷积的原有行为，让模型分别学习左右双向特征表征。
这套架构将紧凑长程记忆、可选择的块内交互能力与双向局部感知能力融合，使GDN架构更加适配块级并行去噪范式。
结语：押注开源端侧Agent路线，harness产品近期上线
当前大模型产业仍由自回归架构主导，推理栈、Agent框架和部署工具链大多围绕其建立，扩散语言要真正进入产业，要体现出自身与现有工具生态的兼容性、开发者迁移成本和规模化交付的稳定性。
扩散智能选择以开源、端侧、智能体能力作为切入点，正是希望降低试用门槛，方便开发者与终端厂商快速开展验证，加速技术落地。
与此同时，扩散智能团队透露其正在研发智能体运行框架（harness）及桌面编程客户端。相关工具尚处于开发阶段，计划于近期上线。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
The author, Wang Xin
Editor, Lee Shui-qing
On October 28th, Shigeki reported that today, Shenzhen ' s LLM manufacturer DiffuSpace publish and open-source proliferation language model Clover-35B-A3B-Preview, using a hybrid specialist structure (MoE), with a total parameter of 35 billion and activating parameter of 3 billion, supported thinking patterns.
The company ' s publish benchmark test results show that Clover-35B-A3B-Preview is ranked first in the scale-regression model Qwen3.6-35B-A3B in six AI agent programming and task assessment missions for tools such as SWE-bench Verified, Terminal-Bench2.1.
HuggingFace：
https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
GitHub：
https://github.com/DiffuSpace/Clover-35B-A3B-Preview
ModelScope：
https://modelscope.cn/models/Diffuspace/Clover-35B-A3B-Preview
Applying for online experience: https://chat.diffuse.ai/
At the same time, proliferation intelligence has introduced the Clover-35B-A3B-Preview decision-making model, which is oriented towards quick, structured diagnostic tasks and supports both stand-alone and joint models, which will be followed by publish.
Unlike the single-point assessment approach of open-source decision-making models such as Jev in TypeSafe, joint models conduct reasoning in a holistic and integrated manner, save time for multiple interactions, allow multiple judgements to cross-reference, circumvent logic conflicts, complete a set of linkage decisions once and achieve faster and more stable reasoning.
In the internal validation of 500 clusters, 2420 links, the joint model answers 95.0 per cent and all 80.4 per cent of the question groups. In contrast, 89.8 per cent and 62.4 per cent, respectively, were the results of the same model when it was independently answered; and Jev ' s historical assessment of the same subject was 63.6 per cent and 15.8 per cent, respectively.
In comparative cases, the five judgements of the joint Clover-35B-A3B-Prepew decision model were all correct, the stand-alone model answered four and the Jev answered two.
The decision model also launched a Chinese chess match with Jeff, with the result that the decision-making model that spreads intelligence prevailed.
The proliferation of intelligence was founded in Shenzhen in 2026 by Prof. Koon Bong Peng, co-director of the NLP Laboratory in Hong Kong, and his Ph.D. Dr. Jinshan San San and Ye Jia Sung. A long-term study of the proliferation language model by the core team was one of the first global teams to explore the route, leading a global breakthrough on the core issues of condition generation, discrete signal modelling and Scaling ' s three DLLM.
As early as 2022, the team initiated the DLLM front-line study, proposing DiffuSeq, which created the technological path of the proliferation model processing sequence to Seq2Seq. In April 2025, it launched the proliferation language model Dream 7B, which to date has accumulated over 2.7 million downloads from Hugo Face.
On 9 October, proliferation intelligence first announced financing, completing two successive rounds of financing, amounting to hundreds of millions of yuan, creating the largest publicly disclosed global record of d LLM financing.
I. 6 Agent first, single GB300 at 680 tokens/s
In terms of benchmarking, Clover-35B-A3B-Preview performed well in the AI agent programming and tool use tasks. In the six Agent evaluations of this comparison, the model achieved the highest scores in the models listed.
Of these, SWE-bench Verified reached 73.6 points, Terminal-Bench 2.1 51.7 points and Overall 65.3 points, exceeding LaDA2.2 flash, Mercury 2.5 and self-regression Qwen 3.6-35B-A3B and Qwen3.5-35B-A3B of the same size.
In many assessments of command compliance, knowledge, logical reasoning and long text understanding, the results reached or exceeded the same scale self-regression model, reaching 83.8, 95.8 and 72.0, respectively, for GPQA Diamond, ZebraLogic and AA-LCR.
In terms of reasoning efficiency, the team tested the internal Agent mission performance records of scenarios such as over-cover code writing, terminal operations and tool call.
With regard to end-side deployment, Clover-35B-A3B-Preview is on a single NVIDIA GB300 GPU at BF16 accuracy, Batch size = 1 to achieve overall production speed of 680 tokens/s.
In addition, on the single NVIDIA H200 GPU, Clover reached 515 tokens/s, up from 365 tokens/s in the DFlash2 programme.
Clover-35B-A3B-Preview achieved the highest scores in six of the nine assessments compared to other representative lightweight, efficient reasoning or end-side deployment models, with advantages over knowledge reasoning, tool use and AI agent programming.
ii. Route planning is 28 steps complete, 49 Agent missions take only 1.52 seconds.
Based on Clover-35B-A3B-Preview ' s ability to reason, use tools and program, users can use the model to do routing, WebMCP missions, simulated desktop environment, AI agent movements, etc.
For example, the user is asked “How long it takes to drive from Pittsburgh CMU to the nearest Wendy's”. Clover identified deviations from the target, gradually entering city, block and mail rewrite search, and then reset and re-routed twice after the page had been stuck, eventually finding the target door shop, and finding a distance of 2.0 km and a distance of about three minutes.
In this comparison, Clover-35B-A3B-Preview completed its mission in 28 steps and 169.9 seconds; Qwen3.6 Flash, which spent 30 steps and 578.3 seconds, remained in a doorway about 4/2 hours from the starting point of 347 kilometres.
Through the web-based tools, Clover-35B-A3B-Preview can continue to perform such tasks as ticketing, course registration, aggregation of amounts, event status and customer information.
In the complete assessment of 49 missions, each of the three attempts, the median time taken by Clover Agent was 1.52 seconds and Gemini 3.6 Flash was 8.86 seconds.
In addition, Clover-35B-A3B-Preview can create private projects and invite members. Under the same AI agent framework, prompt, tools and browser configuration, Clover-35B-A3B-Preview completed its mission in 13 steps, 169.9 seconds, and Qwen 3.6 Flash completed it in 21 steps, 392.3 seconds.
In this case, the user requested to simulate a desktop environment with a single HTML file to support the window drag and to open several small applications.
Clover-35B-A3B-Preview produces analog desktops that include applications such as programmers, window systems, and notebooks, calculators, clocks, weather and settings.
In a demonstration, the user can open the application, enter the text, operate the calculator and adjust the settings.
In the urban low-altitude drone movement demonstration, the model usually requires the interpretation of composite air traffic control messages into tasks such as no-fly zone clearance, site modification, task assignment, no-fly zone radius adjustment and emergency dispatch.
Clover-35B-A3B-Preview produced five correct results after 0.97 seconds of simultaneous processing. Qwen presented the results in 3.78 seconds and misconstrued the UAV5 mapping mission as a medical mission.
Reconstruct the two-way Delta network, balancing the speed of reasoning with performance
Clover-35B-A3B-Preview is able to balance the speed of reasoning with performance, behind which is the optimization of the structure that spreads the intelligent team.
In Clover-35B-A3B-Preview, the proliferation intelligent team completed the redesign of the two-way door-control Delta Network (Bidirective GDN) using a differentiated calculation paradigm for historical memory and block-based noise chains.
The context information that has been generated by the model has been compacted efficiently by GDN regression, and Token in the current block has been able to interact with fine grains with two-way Softmax attention.
At the core, the technical team reads the historical memory as a stand-alone information module, filters the information with Token in the same Softmax operation as the block, and learns the weight of reliance on historical information and current block information in the reasoning process by modeling.
At the same time, the team added a learning block calibration mechanism to mitigate the historical weight dilution caused by the increased number of Tokens in the block.
For local feature extraction, the proliferation intelligence team has also created an independent zero-initialized collage branch on the right side of context. The programme is based on pre-trained pre-employer behaviour, which allows models to learn both directional and directional characterizations.
The architecture combines long-range memory, optional intra-block interactive capabilities with two-way local sensory capabilities, making the GDN structure more compatible with block-level parallels of noise.
Conclusion: A bet on the Agent route on the side of the open source end, with the Harness product coming online in the near future
The current LLM industry is still dominated by a self-regressive structure, around which the reasoning, the Agent framework and the deployment tool chain are mostly built, and where the proliferation of language is needed to truly enter the industry, reflecting its ecological compatibility with existing tools, the cost of developing country migration and the stability of scale delivery.
The diffusion of smart choices using open-source, end-side, AI agent capabilities as entry points is the desire to lower the trial threshold, facilitate rapid validation by developers and end-processors, and accelerate technology landing.
At the same time, the proliferation intelligent team revealed that it was developing a AI agent operating framework (harness) and a desktop programming client. The tools are still in the development stage and are planned to be online in the near future.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.