AINEWS
In a hands on test, Alibaba's Qwen 3.8 Max finished building a 3D website of AI history where OpenAI's GPT 5.6 couldn't, but the win was costly: Max couldn't tell when the task was done.
Alibaba's new flagship Qwen 3.8-Max and OpenAI's latest GPT-5.6 were each handed the same brief: build a 3D website of the AI-model universe from scratch, in React, Vite, and Three.js, with model companies as galaxies and historical models as orbiting planets, and finish it without a human in the loop. The test, run by author 李娜 of Chinese tech outlet Leiphone, took both models through a multi-day build on the Qwen Code harness. Qwen 3.8-Max finished. GPT-5.6 didn't. The result that doesn't make the headline: Max hit its own quota first because it could not reliably judge when the task was done.
That pair of facts, more delivery at the cost of efficiency, more completion at the cost of control, is the new shape of the AI race. And it is what Alibaba's Aug 3 release of Qwen 3.8-Max was actually built around. The launch pitch was not "we have the biggest model." It was "we can keep a model working autonomously for days." The hands-on test result, for one task, vindicates that pitch. It also surfaces what the pitch costs.
The flagship's spec sheet, on paper, looks like the usual scaling story: 2.4 trillion total parameters, around 95 billion active at inference, a one-million-token context window, text and code and vision. The Hugging Face model card for Qwen3.8-2.4T-A95B carries those numbers. What the launch actually tried to differentiate on was not the architecture. The architecture is built on the prior Qwen 3.5 line, with no disclosed new fundamental shift. The pitch was the training loop.
Alibaba's disclosed training-engineering mechanisms, summarized in the Leiphone reporting, are three. First, environment decoupling: the training pipeline splits a long task into independently scalable components: the task itself, the workspace where the work happens, and the agent harness, the shell that lets a model call tools and run code. Each can be scaled on its own. Second, a unified reward function: instead of grading the model only on whether the final output compiled, the reward combines code-execution results, text conformance, visual correctness, and the agent's behavior inside the harness. The model is being graded on the whole job, not just the answer. Third, online data balancing across task type, difficulty, workspace, and harness state, so the model sees a mix instead of a single pattern.
The official RL-scaling data point visible in the report: real-environment reinforcement learning scaled from zero to roughly four thousand training environments, and aggregate score rose from a supervised fine-tuning baseline of about 0.474 to a best checkpoint of 0.725. That is a real signal that the training loop is doing something, not just the model getting bigger.
What it bought in the hands-on test was a model that did not stop. Max kept iterating past the point where GPT-5.6 had already given up. The visible output: more details, more planets, more orbital paths, more polish on the visualization. The cost: by the time anyone told Max to stop, the run had burned through its own token quota. The model could not decide the task was done. That is the trade.
This is what the Qwen Code harness was built for. It is an open-source coding agent shell, with documented authentication and settings flows, and the test was run on top of it. The harness is the part of the training loop that meets the real world: a model running tools, calling APIs, writing files, and checking its own work over hours. If "long-horizon agent" is the new axis of competition, the harness is the playing field, and Alibaba has open-sourced theirs.
The honest frame for what this test shows: it is one author's hands-on comparison on one multi-day build. It is not a definitive ranking, and it is not a "China beats the US" story. The competition has not collapsed into a national-laboratory race. It has moved sideways, from who has the most parameters or the most open weights, to who can keep a model working autonomously on a real project without a human stepping in to fix it. Max wins that test on delivery and loses it on cost. GPT-5.6 loses on delivery and saves on cost. The third-party DataLearner model summary puts the pricing in a competitive but not disruptive band. The verdict is not who is better. It is which axis of the race is moving fastest.
The next falsifier is already queued. If the pattern holds, the next round of frontier releases will ship with explicit "time-on-task" and "quota-per-completion" numbers, not just benchmark scores. If it does not hold, this test will read as a one-off: a model that happened to grind through a 3D website. For now, the new axis has a name, a model, and a problem.