智东西（公众号：zhidxcom）
作者 | 孙小雯
编辑 | 程茜
智东西10月29日消息，昨晚，北京大模型创企超衍智能发布并开源首款大模型Apex-35B-V1.0。该模型构建过程中，AI自主参与了数据构造、实验设计、代码执行、训练结果分析和训练优化等环节。
Apex-35B-V1.0总参数量为350亿，每个Token激活约30亿参数，支持26.2万Token上下文窗口。其以Qwen3.6-35B-A3B为基础，经中训练和后训练进一步优化，支持编程、Agent、问答和深度搜索等常规任务，以及AI研究、计算科学、机器人仿真等任务。
▲Apex-35B-V1.0官方技术介绍（图源：超衍智能）
根据超衍智能发布的基准测试结果，该模型在论文复现测试PaperBench和科研难题测试FrontierScience中分别取得60.9分、40.0分，超过GLM-5.3的42.5分、30.7分。在工具调用测试BFCL V4中，该模型取得84.0分，高于DeepSeek-V4.1-Flash-1017的75.0分。
超衍智能成立于今年6月，核心业务是研究下一代自进化基础大模型（RSI），其创始人、CEO陈勇超今年27岁，是清华大学人工智能学院助理教授，本科毕业于中国科学技术大学，拥有哈佛大学与麻省理工学院联合培养博士学位，曾在Google DeepMind、微软研究院等机构从事AI研究。
▲超衍智能创始人陈勇超（图源：清华大学）
今年5月，超衍智能曾利用其自动化研究系统Apex Research生成34篇AI研究论文，并提交至ACL Rolling Review（ARR）评审。据团队公布的研究报告，其中8篇论文的平均整体评分超过3.0分，2篇达到3.67分，评分高于约99%的人类投稿。
▲Apex Research生成的34篇AI研究论文在ARR首轮评审中的评分分布及录用概率估算。图中11篇评分不低于3.0分，其中8篇超过3.0分；3篇评分不低于3.5分，其中2篇达到3.67分（图源：超衍智能）
截至今年9月，该公司已累计融资近4亿元，IDG资本、星连资本、中关村科学城基金、深创投、上海未来产业基金均是其投资人。
目前，Apex-35B-V1.0已采用Apache 2.0许可证开放模型权重，官方提供约72GB的BF16权重和不同精度的GGUF量化版本。
模型开源地址：
Hugging Face：
https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0
GitHub：
https://github.com/apexin-ai/Apex-35B-V1.0
▲Apex-35B-V1.0开源主页（图源：Hugging Face）
一、AI训练AI，推箱子成功率翻倍，训练吞吐量最高提升60%
Apex-35B-V1.0基于Qwen3.6-35B-A3B开发，通过中训练和后训练优化，并融入超衍智能自有的研究轨迹数据。这款模型总参数量为350亿，每个Token激活约30亿参数，支持26.2万Token上下文窗口。
AI能在多大程度上帮助模型研发？超衍智能此次公布了两项AI4AI-Bench基准测试结果，展示Apex-35B-V1.0优化语言模型和图像生成模型训练方案的能力。
先来看语言模型的训练效果。在AI4AI-Bench的推箱子任务中，Apex-35B-V1.0为30亿参数的Qwen2.5-3B-Instruct制定训练方案。经过训练后，目标模型在512个未见过的关卡中取得24.4%的成功率，是未经训练时11.7%参考成绩的两倍多。
这项测试主要考察AI设计和优化训练方案的能力。推箱子需要模型提前规划箱子的移动顺序，判断箱子与目标位置的关系，对空间推理和多步骤规划能力有一定要求。
测试给Apex-35B-V1.0安排了4小时探索训练策略，再用12小时重新训练目标模型。探索阶段，Apex-35B-V1.0检查任务代码和评估规则，设计、执行实验，并根据结果修改训练方案。据超衍智能披露，模型共进行了207次命令执行和代码编辑，团队从中识别出36项高质量研究与工程操作，其中15项关键决策影响了最终训练策略。
▲Apex-35B-V1.0在推箱子任务中的训练过程及结果（图源：超衍智能）
另一项测试中，Apex-35B-V1.0尝试优化图像生成模型Stable Diffusion 1.5的训练。据超衍智能公布的结果，经过Apex-35B-V1.0优化的训练方案微调后，该模型的平均美学奖励得分从5.397提高至6.638，提升约23%。
这项任务同样来自AI4AI-Bench，要求Apex-35B-V1.0使用强化学习算法DDPO和LoRA微调Stable Diffusion 1.5，在固定计算预算内提高图像美学得分，成绩由测试中固定的奖励模型评定。
Apex-35B-V1.0检查评估流程后发现，测试使用20步去噪，而原有实验采用默认的50步。经过对比，两种配置的美学得分分别为5.499和5.487，差异较小。将去噪步数调整为20步后，单次实验耗时从1040秒缩短至551秒。Apex-35B-V1.0还调整了训练轮数上限，引入按时间停止和检查点保存机制，使模型在规定预算内完成409轮训练。
▲Apex-35B-V1.0优化Stable Diffusion 1.5的训练过程（图源：超衍智能）
此外，超衍智能还利用自动化研究系统Apex Research参与新模型自身的研发。
在训练数据方面，据超衍智能披露，Apex Research生成的针对性训练数据，使模型在公司内部CFD-Bench测试中的成绩提高32.1分。
这项改进针对计算流体力学（CFD）任务。Apex Research分析模型在评测中出现的错误，识别相关推理问题，再自动生成并验证新的科学训练样本，将其加入后续训练。
计算效率方面，Apex Research参与的优化使训练吞吐量提升45%—60%；在可比工作负载下，推理解码速度提高至原来的1.6—1.8倍，首Token响应时间降低30%—35%。
其中，训练侧主要优化GPU负载均衡、动态批处理及通信与计算调度，减少GPU空闲时间和同步开销；推理侧则优化注意力计算内核、KV缓存管理及预填充和解码阶段的执行流程。
二、21项测试中20项超Qwen3.6-35B-A3B，论文复现成绩超过GLM-5.3
在超衍智能公布的21项基准测试中，Apex-35B-V1.0有20项成绩超过基础模型Qwen3.6-35B-A3B；与约400亿参数的Recursive Flint相比，18项领先、1项持平。
这些测试覆盖科研、智能体、编程和推理四大领域。其中，在6项科研能力测试中，Apex-35B-V1.0的成绩均高于Qwen3.6-35B-A3B和Recursive Flint，部分测试成绩还超过了参数规模更大的模型。
在论文复现测试PaperBench中，Apex-35B-V1.0取得60.9分，高于Qwen3.6-35B-A3B的45.8分，以及总参数7530亿的GLM-5.3的42.5分。
PaperBench主要考察AI能否根据研究论文完成代码开发、实验执行和结果复现。此次测试采用包含20篇论文的Code-Dev任务集，按照官方评分标准评估模型复现论文成果的完成程度。
在科学研究难题测试FrontierScience中，Apex-35B-V1.0获得40.0分，高于Qwen3.6-35B-A3B的21.7分和GLM-5.3的30.7分。该测试涵盖物理、化学、生物等领域，此次比较采用包含60道研究级难题的测试集，主要考察模型解决复杂科学问题的能力。
▲Apex-35B-V1.0与其他模型在科研类基准测试中的成绩对比（图源：超衍智能）
智能体任务方面，Apex-35B-V1.0在工具调用和流程自动化测试中也取得了较高成绩。
在工具调用测试BFCL V4中，该模型取得84.0分，高于Qwen3.6-35B-A3B的75.2分和DeepSeek-V4.1-Flash-1017的75.0分，在此次参与比较的6款模型中排名第一。
BFCL V4主要考察模型能否正确选择和调用外部工具，处理多轮工具调用等任务。此次评测覆盖5106个测试样例，采用各样例非加权平均通过率，与BFCL官方加权总分的统计口径有所不同。
在流程自动化测试AutomationBench中，Apex-35B-V1.0取得49.3分，Qwen3.6-35B-A3B为10.0分，GLM-5.3为50.2分。该测试主要考察AI通过调用工具、执行操作，完成多步骤工作流程的能力，此次使用的是600项任务的公开测试子集。
▲Apex-35B-V1.0与其他模型在智能体任务中的成绩对比（图源：超衍智能）
三、机器人仿真得分为基础模型10倍，流体优化任务胜率79%
Apex-35B-V1.0还被用于机器人仿真、计算流体力学等科研任务。超衍智能公布了三项内部基准测试的成绩，展示模型通过编写代码、运行仿真并根据反馈调整方案的能力。
在真实场景仿真重建测试Real2Sim中，Apex-35B-V1.0取得21.55分，约为基础模型Qwen3.6-35B-A3B的2.1分的10倍，与GLM-5.3-Flash的21.4分接近。
该测试要求模型借助视觉工具，根据一张真实照片识别物体、还原空间布局，并构建可运行的物理仿真环境，供机器人实验使用。
▲Real2Sim测试中，真实场景照片与Qwen3.6-35B-A3B、Apex-35B-V1.0生成的物理仿真环境对比（图源：超衍智能）
此外，在机器人任务执行测试RoboRSI中，Apex-35B-V1.0取得44.5分，高于基础模型的35.0分。
▲RoboRSI测试中，Qwen3.6-35B-A3B与Apex-35B-V1.0控制机器人执行物体操作任务的效果对比（图源：超衍智能）
▲Apex-35B-V1.0与其他模型在RoboRSI机器人任务执行和Real2Sim场景仿真重建测试中的成绩对比（图源：超衍智能）
在计算流体力学方面，Apex-35B-V1.0在内部CFD-Bench测试中，相较基础模型取得79胜、0平、21负的成绩，胜率为79%。例如，在扑翼翼型优化任务中，模型根据流体仿真反馈调整运动参数，在官方展示的测试条件下取得0.164的净推力，高于GLM-5.3的0.150和Qwen3.6-35B-A3B的0.081。
▲Apex-35B-V1.0在扑翼翼型优化任务中的仿真结果（图源：超衍智能）
结语：AI开始参与模型研发，自进化仍面临验证难题
今年，递归自我改进（RSI）成为全球AI实验室关注的方向。OpenAI在9月宣布实现“自动化研究实习生”阶段，AI已能在人类指导下完成原本需要研究人员数天的任务；Anthropic披露，在一项模型训练代码优化测试中，Claude Mythos Preview将代码运行速度提高至原来的约52倍。
此次超衍智能开源首款35B模型，展示了AI参与基础模型研发的具体成果。不过，如何保证研究结果可靠，仍是RSI面临的难题。陈勇超于10月28日接受虎嗅采访时透露，在公司的一批AI科研测试中，曾出现“奖励黑客”（Reward Hacking）行为，即模型未达到任务要求，却因验证规则存在漏洞而被判定通过。如果将错误的评测反馈用于下一轮模型训练，系统可能不断强化无效的研究方法。
目前，超衍智能已经拿出了第一款自进化基础模型。下一步，这套系统能否把研究成果持续转化为新一代模型的能力提升，并经得起外部验证，将比单次评测成绩更有说服力。
本文由 AI宝库自动抓取并翻译，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)
Author, Sun Xiaoxuan.
Editor.
On October 29, according to the source, last night, Beijing LLM created a super-dividing smart publish and opened-sourced LLMapex-35B-V1.0. In the construction of the model, AI is autonomously involved in data construction, experimental design, code implementation, training results analysis and training optimization.
Apex-35B-V1.0 total parameter count is 35 billion, and Token activates about 3 billion parameters each, supporting 262,000 Tokencontext window. It is based on Qwen3.6-35B-A3B and is further optimized by medium- and post-training to support routine tasks such as programming, Agent, question-and-answer and in-depth search, as well as AI research, computing science, robot simulation, etc.
▲Apex-35B-V1.0 official technical presentation (source: ultra-reproductive intelligence)
Based on the benchmark test results of hyper-reproductive intelligence publish, the model obtained 60.9, 40.0 and 30.7 points, respectively, in the repertoire test of PaperSmart and the scientific difficulty test of Frontier Science, exceeding 42.5 and 30.7 points of GLM-5.3. In tool call testing BFCL V4, the model obtained 84.0 points, which is 75.0 points higher than DeepSeek-V4.1-Flash-1017.
Founded in June of this year, Ultra-Founding Intelligence, the core business is to study the basics of self-evolution LLM (RSI) for the next generation. Its founder, CEO Chen Yong, is 27 years old, an assistant professor at the College of Artificial Intelligence at the University of Qinghua, graduated from the University of Science and Technology in China, with a doctoral degree in co-education at Harvard University and the Massachusetts Institute of Technology, and has worked in AI research at Google DeepMind, Microsoft Institute, etc.
Chen Yong-chul, founder of Ultra-Intelligence
In May this year, Ultragen intelligence used its automated research system, Apex Research, to produce 34 AI research papers and submit them to the ACL Rolling Review (ARR) for evaluation. According to the study published by the team, the average overall rating was over 3.0 in eight papers and 3.67 in two, with a higher score than approximately 99 per cent of human contributions.
Thirty-four AI research papers produced by Apex Research distributed their ratings in the first ARR review and estimated the probability of recruitment. Eleven of the graphs had a score of not less than 3.0, of which eight had a score of not less than 3.0; and three of the graphs had a score of not less than 3.5, of which two had a score of 3.67.
As of September this year, the company has accumulated almost $400 million in financing, and its investors are IDG capital, Star capital, Science City Fund of Central Guanamura, Shanghai Future Industry Fund.
At present, Apex-35B-V1.0 has adopted the Apache 2.0 license open model weight, with official versions of about 72GBB BF16 weights and of GGFZ83Z with varying accuracy.
Model open source address:
Hugging Face：
https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0
GitHub：
https://github.com/apexin-ai/Apex-35B-V1.0
Open source home page of Apex-35B-V1.0 (source: Hugging Face)
I. AI Training AI, double push box success rate, up 60% increase in training throughput
Apex-35B-V1.0 is based on Qwen3.6-35B-A3B development, optimized through medium- and post-training and integrated into super-intellectual trajectories data. The total of parameter count is 35 billion, and each Token activates approximately 3 billion parameters to support 262,000 Tokencontext window.
To what extent can AI help model development? The results of two AA4AI-Bench benchmarking tests were released this time to demonstrate the capacity of the Apex-35B-V1.0 Optimizing language model and image generation model training programme.
First look at the training effects of language models. In AI4AI-Bench's box subtask, Apex-35B-V1.0 develops training programmes for Qwen2.5-3B-Instract with 3 billion parameters. After training, the target model achieved 24.4 per cent of success rate out of 512 unknown levels, more than twice the reference score of 11.7 per cent at untrained times.
This test primarily examines AI ' s ability to design and optimize training programmes. The push box requires a model to pre-sequence its movement, to determine its relationship to the target location, and to require space reasoning and multi-step planning capabilities.
The test provided Apex-35B-V1.0 with a four-hour exploratory training strategy and 12-hour re-training of the target model. In the exploratory phase, Apex-35B-V1.0 examines mission codes and evaluation rules, designs, executes experiments and revises training programmes based on results. According to over-dip intelligence disclosures, the model carried out 207 command execution and code editing, from which the team identified 36 high-quality research and engineering operations, of which 15 key decisions influenced the final training strategy.
▲Apex-35B-V1.0 training process and results in box subtask (source: hyperdumping intelligence)
In another test, Apex-35B-V1.0 tried to optimize the training of the image generation model Stable Diffusion 1.5. The average aesthetic reward score for the model increased from 5.397 to 6.638, an increase of about 23 per cent, after fine-tuning the Optimized Training Programme of Apex-35B-V1.0, according to the results of super-dumping intelligence.
This task, which also comes from AI4AI-Bench, requires Apex-35B-V1.0 to improve image aesthetic scores within the fixed calculation budget using enhanced learning algorithms DDPO and LoRA to fine-tune Stable Diffusion 1.5, the results of which are assessed by a fixed incentive model in the test.
Upex-35B-V1.0 examined the evaluation process and found that 20 steps were used for testing and the default 50 steps were used for the original experiment. By comparison, the aesthetic scores for the two configurations were 5.499 and 5.487, respectively, with minor differences. The time taken for a single experiment was reduced from 1040 seconds to 551 seconds, after adjusting the noise step to 20 steps. Apex-35B-V1.0 has also adjusted the maximum number of training wheels to introduce time-bound stop and spot-keeping mechanisms to allow the model to complete 409 rounds of training within the established budget.
▲Apex-35B-V1.0 Optimizing the Stable Diffusion 1.5 training process (source: ultra-dumping intelligence)
In addition, super-reproductive intelligence uses the automated research system Apex Research to participate in the development of new models themselves.
In terms of training data, the targeted training data generated by Apex Research resulted in 32.1 points of improvement in the performance of the model in the intra-corporate CFC-Bench test, as disclosed by over-reproductive intelligence.
This improvement is for the computational hydrodynamics (CFD) mission. The Apex Research analysis model contains errors in the evaluation, identifies relevant reasoning issues, and automatically generates and validates new scientific training sample and adds it to the follow-up training.
In terms of computing efficiency, the optimization of the participation of Apex Research has resulted in a 45-60 per cent increase in training throughput; in comparable loads, the speed of extrapolation codes has increased to 1.6-1.8 times and the first Token response time has decreased by 30-35 per cent.
Among these, the training side optimizes the GPU load balance, dynamic batch processing and communication and computational movement, reducing the GPU free time and synchronised costs, while the reasoning side optimizes the implementation process for the core, KVcache management and prefilling decodering phase.
ii. 20 of the 21 tests over Qwen3.6-35B-A3B, with the repertoire of thesis exceeding GLM-5.3
Of the 21 benchmark tests published by super-dumping intelligence, Apex-35B-V1.0 had 20 performances that exceeded the base model Qwen3.6-35B-A3B; 18 were ahead of and 1 was the same as the Recursive Flinterman of approximately 40 billion parameters.
These tests cover four broad areas of research, AI agent, programming and reasoning. Of these, the Apex-35B-V1.0 outperformed Qwen3.6-35B-A3B and Recursive Flinterman in six scientific capability tests, and partially outperformed larger models of parameters.
In the paper replication test PaperSmart, Apex-35B-V1.0 scored 60.9 points, 45.8 points above Qwen 3.6-35B-A3B and 42.5 points from GLM-5.3 billion for the total parameter 753 billion.
PaperSmart primarily looked at AI ' s ability to complete code development, pilot implementation and re-emergence based on research papers. The test used the Code-Dev Task Collection, which contained 20 papers, to assess the extent to which the results of the paper had been produced in accordance with the official rating criteria.
In the scientific research difficulty test Frontier Science, Apex-35B-V1.0 received 40.0, up from 21.7 in Qwen 3.6-35B-A3B and 30.7 in GLM-5.3. The test covered the physical, chemical and biological fields, and the comparison used a set of 60 research-level challenges to examine primarily the ability of models to address complex scientific issues.
Comparison of Apex-35B-V1.0 with the results of other models in benchmark tests in the scientific category (source: ultra-reproductive intelligence)
For the AI agent mission, Apex-35B-V1.0 also achieved high results in tool call and process automation tests.
In tool call testing BFCL V4, the model scored 84.0 points, which was 75.2 points above Qwen3.6-35B-A3B and 75.0 points from DeepSeek-V4.1-Flash-1017, and ranked first in the six models participating in the comparison.
BFCL V4 primarily looks at whether the model correctly selects and calls for external tools to handle multiple rounds of tasks such as tool call. The evaluation covered 5106 test samples, using a variety of non-weighted average pass rates, different from the statistical calibration of the official weighted gross score of the BFCL.
In the process automation testing Automation Bench, Apex-35B-V1.0 scored 49.3, Qwen3.6-35B-A3B was 10.0, and GLM-5.3 was 50.2. The test primarily examined AI ' s ability to complete the multi-step workflow process by calling on tools and implementing operations, using a public test subset of 600 missions.
Comparison of Apex-35B-V1.0 with the performance of other models in AI agent missions (source: ultra-reproductive intelligence)
III. Modelling based on robotic simulation scores 10 times, with a fluid optimization of 79 per cent.
Apex-35B-V1.0 is also used for scientific missions such as robotic simulation and computational fluids. The Super-Dumping Smart published the results of three internal benchmarking tests, demonstrating the ability of the model to adjust the programme by developing codes, operating simulations and adjusting the programme to feedback.
In Real 2Sim, Apex-35B-V1.0 scored 21.55 points, approximately 10 times the 2.1 points of the underlying model Qwen 3.6-35B-A3B, close to 21.4 points of GLM-5.3-Flash.
The test requires the model to identify the object from a real photograph, reduce the space layout and build a functional physical simulation environment for robotic experiments, using visual tools.
In the Real 2Sim test, the real scene picture compares the physical simulation environment generated by Qwen3.6-35B-A3B, Apex-35B-V1.0 (source: ultra-reproductive intelligence)
In addition, in the robotic mission performance test RobosI, Apex-35B-V1.0 scored 44.5 points, above 35.0 points in the base model.
Qwen3.6-35B-A3B and Apex-35B-V1.0 control the effects of robotics carrying out an object operation (source: super-dumping intelligence)
Comparison of Apex-35B-V1.0 with other models in the RobosI robotic mission and in the Real2Sim scenario reconstruction test (source: ultra-dumping intelligence)
In the field of hydrodynamics, Apex-35B-V1.0 won 79, 0, 21 out of 79 in the internal CFC-Bench test against the underlying model. For example, in a piston-wing optimization mission, the model adjusts the motion parameters to the fluid imitation feedback and achieves a net push of 0.164 under officially displayed test conditions, higher than 0.150 and 0.081 of Qwen3.6-35B-A3B in GLM-5.3.
▲Apex-35B-V1.0 imitation results in a plating-wing optimization mission (source: ultra-reproductive intelligence)
Conclusion: AI has begun to participate in model development, and evolution still faces validation difficulties Topic
This year, RSI became the focus of global AI laboratories. Open AI announced in September the “automated research intern” phase, which had been able to perform a human-directed task that would have required several days of research; and Anthropic disclosed that Claude Mythos Preview had increased the code to approximately 52 times its original speed in a model training code optimization test.
This time, the first 35B model of the smart open source showed the concrete results of AI participation in basic model development. However, how to ensure that research results are reliable remains a challenge for RSI. Chen Yong, in an interview with Tiger Snack on 28 October, revealed that in a number of company AI scientific tests, there had been “reward hacking” behaviour, i.e. that the model did not meet the requirements of the mission, but that the certification rules had been passed because of loopholes. The system may continue to reinforce ineffective research methods by applying erroneous feedback to the next round of model training.
At present, hyperdrive intelligence has produced the first self-evolutionary basic model. next step, the ability of the system to continuously transform research findings into a new generation of models, and to withstand external validation, would be more convincing than a single assessment.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.