Alibaba Simply Gave Away Its Finest AI Mannequin For Free, Virtually Matching Claude and ChatGPT – Decrypt

Alibaba launched Qwen3.8-Max on Monday, calling it essentially the most succesful mannequin it has ever constructed. The weights land on Hugging Face and ModelScope subsequent week—the primary time the corporate has given away a mannequin at Max scale. 

The specs are massive: 2.4 trillion parameters whole, with 95 billion switched on at any second. Parameters are the variety of dials a mannequin is ready to deal with. 

That is extraordinarily essential for effectivity because it means it might not require too a lot sources to run. Consider it as an enormous library the place solely the related shelf lights up for every query. Small companies and labs with ok {hardware} are actually capable of run a state-of-the-art mannequin with out spending the identical as an enormous datacenter. 

Alibaba’s launch submit skips the standard benchmark-chasing narrative and leans on endurance as a substitute. The mannequin spent 16 days constructing a coding instrument by itself—265 commits, 127 pull requests, 151 points, no human touching the keyboard. It spent 5 days reproducing a analysis paper it had by no means seen code for, then beat the paper’s personal outcomes by 2.7 factors. In a 24-hour machine studying contest, it completed forward of 458 of 526 human groups. 

Constructed to run inside a rival’s instruments 

Qwen3.8-Max ships with directions for Claude Code and Codex, the coding instruments made by Anthropic and OpenAI. Alibaba’s API speaks each firms’ protocols. Most of its coding benchmarks have been run inside Claude Code. 

And people benchmarks do not flatter it. Throughout 31 textual content assessments, Anthropic’s Fable 5 takes 15 first place spots, OpenAI’s GPT-5.6 Sol takes 9, Qwen takes seven. On the 12 coding assessments, Qwen wins precisely one. Nevertheless, by way of intelligence prices, this mannequin is extraordinarily low cost and environment friendly, which implies that even when it requires extra iterations or reasoning, the price of getting the job performed will likely be a lot decrease, practically 30% of what Claude Fable 5 expenses. 

That mentioned, flip to multimodal work—paperwork, video, spatial reasoning—and the rankings invert. Qwen leads most of that desk. 

There’s additionally a reversal by way of enterprise technique. In April, Alibaba killed the free tier of Qwen Code, with the workforce drifting towards closed, paid fashions after management departures. Our evaluation of Qwen 3.7 Max famous the Plus model can be open whereas Max stayed locked behind the API. 

That door is now open, and the timing is not an accident. Chinese language open-weight fashions went from beneath 2% of tokens on OpenRouter in late 2024 to roughly 61% by mid-2026. Qwen handed Meta’s Llama as essentially the most self-hosted mannequin on the earth. 

In the meantime Washington restricted Fable 5 and Mythos 5 beneath export controls in June, and Beijing is reportedly weighing limits of its personal on Chinese language fashions going abroad. 

So Alibaba is shedding on paper and successful on distribution. For those who can obtain one thing that comes shut without cost, second place is a advantageous place to be.