Seven PhD students trained a 7B model — and left every log in
Beijing's frontier labs get most of the attention, but the most instructive Chinese AI release this week came from a university project with a student-sized budget and an agent-sized workforce.
Seven PhD students at Beijing's Zhongguancun Academy trained a 7B-parameter language model from scratch in one summer, then published everything — code, data recipe, logs, and every intermediate checkpoint. The model is ZGCM-1, a 7.39B dense model released under MIT with a 256K-token context window, aimed at mathematical reasoning and tool-assisted search. According to QbitAI, which reported on the project, the team of seven split across AI, networking, chemistry, biology and cybersecurity backgrounds; they built a small prototype over a few days to convince their advisors to hand over more compute. On the technical report's numbers, ZGCM-1 scores 97.13% on MATH-500, 75.00% on AIME 2026 and 70.42% on HMMT 2025, taking the best average rank across 14 reasoning benchmarks among the seven 7B–8B models compared. Its agentic results — 63.09% on WebWalkerQA and 19.43% on BrowseComp — are where the "small model, think longer" bet paid off, with the team reporting some search and tool-use scores in the same range as much larger models such as Qwen3-235B-A22B and GLM-5.1.
The part other labs should read is the workforce, not the weights. Data at trillion-token scale, cluster operations, experiment monitoring and evaluation are normally separate specialist teams. Seven people could not cover them, so they ran hundreds of agents in sub-teams for data, experiments and evals, wired into a forum-style task board where work gets claimed, reported and checked. Their own ZGent platform accumulated meeting notes, verified scripts and debugging tricks as reusable skills for later agents. Then they graded the experiment honestly: across 11 task categories on an L1–L5 autonomy scale, monitoring and deployment reached L4 — agents plan, execute and adjust from feedback — while model architecture and algorithm design stayed at L2, still mostly helping humans implement what humans decided.
That honesty is what the release is really for. At one point loss kept falling while the model's abilities got worse; the curve showed nothing, and the cause turned out to be drift in the underlying data shards and shuffling, so the actual recipe fed to training wasn't the recipe on paper. The fix was a capability-tracking layer — an internal ACE evaluation of 18 categories, 183 items and 2,503 probes that runs a diagnosis round in two to three minutes. Efficiency work ran alongside it: gated sliding-window plus global attention, context extended in stages from 16K to 64K to 256K for roughly 3.94x throughput versus full attention at 256K with about one-sixth the KV cache, and Muon with FP8 training — using delayed scaling and TWEO to keep extreme activations from wrecking low-precision runs — for an estimated 4.2x improvement in time-to-loss on 16K pretraining. SFT gave its own surprises: tightening quality filters cut samples by about 44.9% and still raised overall eval scores, while too much long chain-of-thought data hurt instruction following.
The publication record is the real contribution. Final weights alone let people use a model; stage checkpoints, training code, data recipes and logs let them see how capability formed and test what each change actually did — exactly the details the students said were missing from the tech reports they kept reading. The team is already starting work at 400B and 500B scale. Whether the method and the agent workforce survive the jump is the next open question, and this time they plan to publish that too. It lands in the same week a 7B open model was beating rivals four times its size locally — K2 Horizon 7B — which makes small, fully-documented models look less like a student exercise and more like a competing research program.
Seven students, hundreds of agents, one summer: is "AI4AI" now the only realistic way for an academic lab to train a frontier-adjacent model? Tell us in the comments.
Sources: ZGCM-1 technical report (GitHub)
