Seven Doctoral Students Train 7B Model ZGCM-1 in Three Months and Release Full Training Record

In general evaluations, ZGCM-1 performed close to same-scale models such as Qwen3-8B. On some mathematical reasoning and search evaluations, it was compared with larger models including Qwen3-235B-A22B and GLM-5.1. The report said the more unusual question was how a seven-person team handled data, training, cluster and evaluation work that at large companies often involves hundreds of people.

The seven students came from different fields: two in artificial intelligence, two in networking, and three in chemistry, biology and cybersecurity. They had mostly fine-tuned existing models. After reading many technical reports, they still wanted to understand how data was selected, how training failures were diagnosed, and how much work phrases such as data cleaning represented. The idea began during a hotpot meal. They initially considered using a 7B model to challenge a model such as Sonnet by letting it think longer and learn to search and use tools. Teachers were skeptical but provided some computing power. The team reproduced existing projects and iterated on a small model, quickly running through the full pipeline, which led to more resources.

When they began training the 7B model, they found they still did not understand many things. Concepts they had seen in papers became obstacles when they had to modify code or explain experimental phenomena. They asked agents to research and explain basic principles, then returned to code and experiments to verify the explanations. Because seven people could not handle trillions of tokens from different sources, uneven quality, cleaning, deduplication, format checks and distribution checks, they built a group of hundreds of agents. The agents were divided into subteams for data, experiments and evaluation, and the team built a forum-like task publishing and reporting system so tasks could be claimed, progress seen, results checked and agent work evaluated. In data processing, agents wrote cleaning scripts based on human quality requirements, checked whether data distributions met standards and automatically adjusted rules and thresholds in a closed loop. In experiments, they could submit tasks, monitor logs, locate faults, adjust configurations, discover storage and data transfer bottlenecks and optimize the training framework's I/O to improve speed. To improve iteration and communication among agents, the team used its ZGent platform to turn meeting discussions, research decisions and plans into shared context. Validated scripts, workflows and debugging experience were stored as reusable Skills, so agents that joined later could build on earlier experience.

The team later categorized the research and development process into 11 types of tasks and asked core participants to rate agent autonomy from L1 to L5. Experimental monitoring and deployment reached L4, meaning that after humans set goals and constraints, agents could plan, execute and continue adjusting based on feedback. Model architecture and learning algorithm design remained at L2, where agents mainly helped implement existing schemes and run preset experiments. Decisions about what to research and which key designs to choose still required humans.

Training problems did not disappear with the agent team. In one run, the loss continued to fall, but the model's capabilities suddenly regressed. The curve did not explain what had happened. The team traced the problem to data sharding and shuffling: the actual proportion of data fed into training fluctuated, so the model did not always receive the recipe they thought it was receiving. After that, they saved intermediate checkpoints more frequently and tracked changes in knowledge, mathematics, code and reasoning. They built the ACE atomic capability evaluation system, which divides capabilities into 18 categories and 183 items and uses 2,503 probes. In their internal distributed evaluation system, one diagnostic cycle took about two to three minutes. At one point during training, they asked the model to write a song, and it did. That night they returned to the hotpot restaurant where the idea began and celebrated. Their understanding of research efficiency also changed. Every idea consumed GPU hours, so they had to ask how to get better results from the same compute or test more hypotheses. The team used an architecture combining local attention and global attention, extending context from 16K to 64K and then 256K during training. At 256K context, the design delivered about 3.94 times the throughput of a full-attention approach and reduced KV cache usage to about one-sixth. They also combined Muon and FP8, used delayed scaling and TWEO to suppress extreme activation values and improve stability in low-precision training. In 16K pretraining, the full scheme reached the same loss with about 4.2 times the efficiency of a standard BF16/AdamW baseline.

After the model could hold information, it still needed to know what to do next. Which result to read after a search, how to recover after a tool failure and whether to change a plan when new evidence appeared all require continuous judgment over long-horizon tasks. The team reorganized interaction trajectories into states and actions, letting the model learn the next decision based on the current task, historical interactions and the latest feedback. Training the model from scratch also produced practical findings that are difficult to obtain from papers and technical reports alone. In the SFT stage, stricter quality filtering reduced samples by about 44.9 percent, while overall evaluation performance improved, although not every capability benefited equally. A proportion of long chain-of-thought data that was too high harmed instruction following. Agent data also could not be trained separately from general capabilities. The team became more concerned with the conditions under which a method worked and the cost it required.

After the final evaluation, the team had a clearer picture of where the model stood. ZGCM-1 led same-scale models in 14 reasoning evaluations, according to QuantumBit, and results on some search and tool-calling tasks showed the potential of smaller models. The team released weights, stage weights, intermediate checkpoints, training code, data recipes and logs. They wanted others to trace how capabilities formed and to check what a change actually brought. Details they had once wanted to see when reading technical reports were kept in their own project as much as possible. The project was guided by He Jiyan and involved faculty and students from the AI Core Division, Software Intelligence Institute and Future Laboratory at Beijing Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence. They have now begun trying 400B and 500B scale models. How much of the methods accumulated at the 7B stage will remain useful at larger scale, and how much new work hundreds of agents can take on, are questions they plan to continue answering through experiments and to record and share.

Editor's Summary

Seven doctoral students at Beijing Zhongguancun Academy trained ZGCM-1, a 7B model, from scratch in three months with a team of hundreds of AI agents handling data, experiments and evaluation. They released weights, code, data recipes, checkpoints and logs, and rated agents at L4 autonomy in monitoring and deployment but L2 in architecture and algorithm design. The team has begun testing 400B and 500B scale models.