Alibaba releases Qwen4-Max: a 2.4-trillion-parameter model with a million-token context, scoring 52.69 on Work360 versus 47.74
Synopsis
Alibaba's Qwen team released Qwen4-Max, built on the new Qwen4 architecture with a 2.4-trillion-parameter backbone, 71B parameters activated per token, native vision and a one-million-token context window. It scores 52.69 on the team's Work360 industry-task benchmark, up from 47.74 for Qwen3.8-Max, uses fewer tokens at matched effort levels on Work360 and Terminal-Bench 4.0, and carries lower API prices. The release also documents pre-training, long-running agent and safety evaluations.
Interpretation
Qwen4-Max moves the previous generation's capability boundary forward as a whole: it is built on the new Qwen4 architecture with a 2.4-trillion-parameter backbone, 71B parameters activated per token and an additional 328B N-gram Embedding parameters, with native vision and a one-million-token context window. The previous generation, Qwen3.8-Max, is the comparison baseline in these evaluations, and the team says Qwen4-Max is stronger on nearly every capability it evaluated. On the team's Work360 industry-task benchmark, Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max; Work360 organizes tasks across 26 Tier-1 domains, 182 Tier-2 subdomains and 721 Tier-3 specialized fields.
It uses fewer tokens and costs less on the same tasks: the team reports token usage dropping by about 24% and 35% at Xhigh and High effort on Work360, and by about 48% and 49% on Terminal-Bench 4.0. API prices are lower than for Qwen3.8-Max, with particularly large reductions in cached-input pricing. In the Work360 evaluation, average cost per task falls from $4.43 to $1.67, a 62% reduction.
On long tasks it delivers usable results in stages so users can step in and correct course, and the team open-sourced the Qwen-Live-Delivery-Plugin. Previously users had to wait until a long-running task finished before seeing results, and an agent's mid-course mistakes were hard to undo. The mechanism has the agent deliver a usable first version early, then expand and improve it through successive releases based on execution results and user requirements.
In materials discovery, Qwen4 ran autonomously for 48 continuous hours on Deep Principle's Mira dry-and-wet-lab platform, starting from a 2.7-million-molecule candidate space and, through four iterations of a design, synthesis-and-testing, active-learning loop, producing a physically synthesized, experimentally validated new coolant molecule. The path closes the loop from digital design to wet-lab synthesis and testing rather than stopping at molecular structures on a screen. The workflow covered structural screening, MPA property prediction, first-principles calculations, retrosynthetic analysis and experiment planning, with an automated wet lab preparing reagents and carrying out synthesis, separation, purification and testing.
Perspective
This report covers the Qwen4-Max model card released by the Qwen team, including architecture, benchmark scores, industry cases and safety evaluations; these results correspond to the team's own or cited evaluation settings, and the industry cases are showcase scenarios. The team says smaller open-source models in the Qwen4 family, trained on the same lineage and techniques, are coming.
The team reports that pre-training at the 2.4-trillion-parameter scale showed steadily decreasing loss with no loss spikes, attributed mainly to architecture and training-optimization improvements; in the base-model comparison, GLM-5.3 and Kimi-K3.1-Pro are post-trained checkpoints, which may affect comparisons with base models. On safety, the team says internal expert review found no capabilities posing large-scale cybersecurity risks and no risk of the model translating relevant knowledge into a significant uplift in a user's capabilities.
