Chaoyan Intelligence Open-Sources 35B Model Apex-35B-V1.0, with AI Participating in Its Own R&D

Zhidongxi October 29 news: last night, Beijing large-model startup Chaoyan Intelligence released and open-sourced its first large model, Apex-35B-V1.0. During the model's construction, AI autonomously participated in data construction, experiment design, code execution, analysis of training results, and training optimization.

Apex-35B-V1.0 has a total parameter count of 35 billion, activates about 3 billion parameters per token, and supports a 262,000-token context window. It is based on Qwen3.6-35B-A3B and further optimized through mid-training and post-training, supporting routine tasks such as coding, agents, question answering, and deep search, as well as tasks such as AI research, computational science, and robot simulation.

▲Official technical introduction of Apex-35B-V1.0 (Image source: Chaoyan Intelligence)

According to benchmark results released by Chaoyan Intelligence, the model scored 60.9 and 40.0 on the paper reproduction test PaperBench and the scientific research challenge test FrontierScience, respectively, surpassing GLM-5.3's 42.5 and 30.7. On the tool-calling test BFCL V4, the model scored 84.0, higher than DeepSeek-V4.1-Flash-1017's 75.0.

Chaoyan Intelligence was founded in June this year, and its core business is researching next-generation self-evolving foundation large models (RSI). Its founder and CEO, Chen Yongchao, is 27 this year and is an assistant professor at Tsinghua University's Institute for AI. He graduated with a bachelor's degree from the University of Science and Technology of China and holds a joint doctoral degree from Harvard University and MIT, and has conducted AI research at institutions including Google DeepMind and Microsoft Research.

▲Chaoyan Intelligence founder Chen Yongchao (Image source: Tsinghua University)

In May this year, Chaoyan Intelligence used its automated research system Apex Research to generate 34 AI research papers and submit them to ACL Rolling Review (ARR) for review. According to the research report published by the team, the average overall score of 8 of the papers exceeded 3.0, and 2 reached 3.67, higher than about 99% of human submissions.

▲Score distribution and acceptance probability estimates of the 34 AI research papers generated by Apex Research in the first round of ARR review. In the figure, 11 papers scored no lower than 3.0, of which 8 exceeded 3.0; 3 papers scored no lower than 3.5, of which 2 reached 3.67 (Image source: Chaoyan Intelligence)

As of September this year, the company had raised nearly 400 million yuan in cumulative funding, with IDG Capital, Xinglian Capital, Zhongguancun Science City Fund, Shenzhen Capital Group, and Shanghai Future Industry Fund all as its investors.

Currently, Apex-35B-V1.0 has released model weights under the Apache 2.0 license, and the official release provides about 72GB of BF16 weights and GGUF quantized versions of different precisions.

Model open-source addresses:

Hugging Face:

GitHub:

▲Apex-35B-V1.0 open-source homepage (Image source: Hugging Face)

Apex-35B-V1.0 is developed based on Qwen3.6-35B-A3B, optimized through mid-training and post-training, and incorporates Chaoyan Intelligence's own research trajectory data. This model has a total parameter count of 35 billion, activates about 3 billion parameters per token, and supports a 262,000-token context window.

To what extent can AI help model R&D? Chaoyan Intelligence this time released two AI4AI-Bench benchmark results, demonstrating Apex-35B-V1.0's ability to optimize training plans for language models and image generation models.

First, look at the training effect on the language model. In the Sokoban task of AI4AI-Bench, Apex-35B-V1.0 formulated a training plan for the 3-billion-parameter Qwen2.5-3B-Instruct. After training, the target model achieved a 24.4% success rate on 512 unseen levels, more than double the reference score of 11.7% without training.

The test gave Apex-35B-V1.0 4 hours to explore training strategies, then 12 hours to retrain the target model. In the exploration phase, Apex-35B-V1.0 inspected the task code and evaluation rules, designed and executed experiments, and revised the training plan based on the results. According to disclosures by Chaoyan Intelligence, the model carried out 207 command executions and code edits in total, and the team identified 36 high-quality research and engineering operations from them, of which 15 key decisions influenced the final training strategy.

▲Training process and results of Apex-35B-V1.0 in the Sokoban task (Image source: Chaoyan Intelligence)

In another test, Apex-35B-V1.0 attempted to optimize the training of the image generation model Stable Diffusion 1.5. According to results released by Chaoyan Intelligence, after fine-tuning with the training plan optimized by Apex-35B-V1.0, the model's average aesthetic reward score rose from 5.397 to 6.638, an increase of about 23%.

This task also comes from AI4AI-Bench, requiring Apex-35B-V1.0 to use the reinforcement learning algorithms DDPO and LoRA to fine-tune Stable Diffusion 1.5 and improve the image aesthetic score within a fixed compute budget, with the score judged by a fixed reward model in the test.

After examining and evaluating the process, Apex-35B-V1.0 found that the test used 20 denoising steps, whereas the original experiment used the default 50 steps. In comparison, the aesthetic scores of the two configurations were 5.499 and 5.487, respectively, a small difference. After adjusting the number of denoising steps to 20, the time required for a single experiment dropped from 1040 seconds to 551 seconds. Apex-35B-V1.0 also adjusted the upper limit on training epochs and introduced time-based stopping and checkpoint-saving mechanisms, enabling the model to complete 409 epochs of training within the specified budget.

▲Apex-35B-V1.0 optimizes the training process of Stable Diffusion 1.5 (Image source: Chaoyan Intelligence)

In addition, Chaoyan Intelligence also used its automated research system Apex Research to participate in the R&D of the new model itself.

Regarding training data, according to disclosures by Chaoyan Intelligence, targeted training data generated by Apex Research raised the model's score on the company's internal CFD-Bench test by 32.1 points.

This improvement targets computational fluid dynamics (CFD) tasks. Apex Research analyzes the errors the model makes in evaluations, identifies the relevant reasoning problems, and then automatically generates and verifies new scientific training samples, adding them to subsequent training.

In terms of computational efficiency, optimizations involving Apex Research increased training throughput by 45%–60%; under comparable workloads, inference decoding speed rose to 1.6–1.8 times the original, and first-token response time fell by 30%–35%.

Among these, the training side mainly optimized GPU load balancing, dynamic batching, and communication-computation scheduling, reducing GPU idle time and synchronization overhead; the inference side optimized attention computation kernels, KV cache management, and the execution flow of the prefill and decode stages.

II. Exceeds Qwen3.6-35B-A3B on 20 of 21 tests, paper reproduction score surpasses GLM-5.3

Among the 21 benchmark tests published by Chaoyan Intelligence, Apex-35B-V1.0 scored higher than the base model Qwen3.6-35B-A3B on 20; compared with the roughly 40-billion-parameter Recursive Flint, it led on 18 and tied on 1.

These tests cover four major domains: scientific research, agents, coding, and reasoning. Among them, in 6 scientific research capability tests, Apex-35B-V1.0 scored higher than both Qwen3.6-35B-A3B and Recursive Flint, and on some tests it even surpassed models with larger parameter counts.

In the paper reproduction test PaperBench, Apex-35B-V1.0 scored 60.9, higher than Qwen3.6-35B-A3B's 45.8 and the 42.5 of GLM-5.3, which has 753 billion total parameters.

PaperBench mainly assesses whether AI can complete code development, experiment execution, and result reproduction based on research papers. This test used the Code-Dev task set containing 20 papers, evaluating the degree to which the model reproduces the paper's results according to the official scoring criteria.

In the scientific research problem test FrontierScience, Apex-35B-V1.0 scored 40.0, higher than Qwen3.6-35B-A3B's 21.7 and GLM-5.3's 30.7. This test covers fields such as physics, chemistry, and biology; this comparison used a test set containing 60 research-level difficult problems, mainly assessing the model's ability to solve complex scientific problems.

▲Comparison of Apex-35B-V1.0 and other models on scientific research benchmarks (Image source: Chaoyan Intelligence)

In terms of agent tasks, Apex-35B-V1.0 also achieved relatively high scores in tool-calling and workflow automation tests.

In the tool-calling test BFCL V4, the model scored 84.0, higher than Qwen3.6-35B-A3B's 75.2 and DeepSeek-V4.1-Flash-1017's 75.0, ranking first among the 6 models compared in this round.

BFCL V4 mainly assesses whether the model can correctly select and call external tools and handle tasks such as multi-turn tool calls. This evaluation covered 5,106 test samples and used the unweighted average pass rate across samples, which differs from the statistical basis of BFCL's official weighted total score.

In the workflow automation test AutomationBench, Apex-35B-V1.0 scored 49.3, Qwen3.6-35B-A3B scored 10.0, and GLM-5.3 scored 50.2. This test mainly assesses AI's ability to complete multi-step workflows by calling tools and executing operations; this round used a public test subset of 600 tasks.

▲Comparison of Apex-35B-V1.0 and other models on agent tasks (Image source: Chaoyan Intelligence)

III. Robot simulation score is 10 times that of the base model, win rate 79% on fluid optimization tasks

Apex-35B-V1.0 was also used for scientific research tasks such as robot simulation and computational fluid dynamics. Chaoyan Intelligence published results on three internal benchmarks, demonstrating the model's ability to write code, run simulations, and adjust its approach based on feedback.

In the real-world scene simulation reconstruction test Real2Sim, Apex-35B-V1.0 scored 21.55, about 10 times the 2.1 of the base model Qwen3.6-35B-A3B, and close to GLM-5.3-Flash's 21.4.

This test requires the model to use visual tools to identify objects from a real photograph, reconstruct the spatial layout, and build a runnable physical simulation environment for robot experiments.

▲In the Real2Sim test, comparison of a real-scene photograph with the physical simulation environments generated by Qwen3.6-35B-A3B and Apex-35B-V1.0 (Image source: Chaoyan Intelligence)

In addition, in the robot task execution test RoboRSI, Apex-35B-V1.0 scored 44.5, higher than the base model's 35.0.

▲In the RoboRSI test, comparison of the effects of Qwen3.6-35B-A3B and Apex-35B-V1.0 controlling a robot to perform object manipulation tasks (Image source: Chaoyan Intelligence)

▲Comparison of Apex-35B-V1.0 and other models on the RoboRSI robot task execution and Real2Sim scene simulation reconstruction tests (Image source: Chaoyan Intelligence)

In computational fluid dynamics, Apex-35B-V1.0 achieved 79 wins, 0 ties, and 21 losses against the base model in the internal CFD-Bench test, for a win rate of 79%. For example, in a flapping-wing airfoil optimization task, the model adjusted motion parameters based on fluid simulation feedback and achieved a net thrust of 0.164 under the test conditions officially presented, higher than GLM-5.3's 0.150 and Qwen3.6-35B-A3B's 0.081.

▲Simulation results of Apex-35B-V1.0 in the flapping-wing airfoil optimization task (Image source: Chaoyan Intelligence)

This year, recursive self-improvement (RSI) has become a direction of interest for global AI labs. OpenAI announced in September that it had reached the "automated research intern" stage, where AI can complete tasks that originally took researchers days under human guidance; Anthropic disclosed that in a model training code optimization test, Claude Mythos Preview increased code running speed to about 52 times the original.

This time, Chaoyan Intelligence open-sourced its first 35B model, demonstrating concrete results of AI participating in foundation model R&D. However, how to ensure the reliability of research results remains a difficult problem for RSI. Chen Yongchao revealed in an interview with Huxiu on October 28 that in a batch of the company's AI research tests, "reward hacking" behavior had occurred, that is, the model did not meet the task requirements but was judged as passing because of loopholes in the verification rules. If erroneous evaluation feedback is used for the next round of model training, the system may continuously reinforce ineffective research methods.

At present, Chaoyan Intelligence has produced its first self-evolving foundation model. The next step is whether this system can continuously translate research results into capability improvements in new-generation models and withstand external verification, which will be more convincing than a single evaluation score.