Qwen 3.7 Max Preview: What Alibaba's New AI Will get Proper and The place It Falls Brief – Decrypt

Alibaba is taking pictures out AI fashions like loopy, and so they’re now way more highly effective than ever with the Qwen 3.7 household out for testing. This week, two new fashions quietly appeared on Enviornment AI’s leaderboard: Qwen 3.7 Max-Preview and Qwen 3.7-Plus-Preview. The fashions are, after all, the appetizer for the Alibaba Cloud Summit 2026.

This is similar playbook Alibaba ran with Qwen 3.6 Max in April. Validate first. Market later. It is a smarter transfer than it appears to be like—Enviornment AI makes use of blind, crowd-sourced comparisons, so the rankings mirror what actual customers truly want, not what a benchmark press launch says.

The outcomes held up. As Decrypt lined when Qwen 3.6 Max dropped, Alibaba has been quietly tightening the hole with Western frontier labs for months. Qwen 3.7 Max-Preview lands at #13 total in Textual content Enviornment, rating seventh in math, ninth in expert-level prompts, and ninth in software program and IT. That makes Alibaba the sixth-ranked AI lab globally in textual content, and fifth in imaginative and prescient capabilities.

The open-source query issues right here. Alibaba killed the free tier of Qwen Code final month and has been shifting its greatest fashions behind a paywall. Qwen 3.7 follows the identical logic: Plus will get open-sourced, Max stays proprietary. The official Qwen 3.7 weblog put up confirms this instantly. Builders who need the perfect Qwen might be paying for it.

That mentioned, the perfect small, open-source agentic coding fashions for native inference are based mostly on Qwen, and this new household guarantees to enhance on what made 3.6 so widespread amongst AI fans.

Each fashions (Plus and Max) are at present locked into deep pondering mode, with net search and code interpreter disabled. It is a preview. The complete launch was anticipated on the Cloud Summit on Might 20.

We ran a fast take a look at on Qwen 3.7 Max to see the way it stacks in opposition to one other Chinese language mannequin, Xiaomi Mimo, which carried out extraordinarily effectively. Here’s what we discovered.

Artistic Writing

We ran Qwen 3.7 Max on the identical immediate we used for MiMo-V2-Professional: a time journey story constructed across the protagonist’s cultural background, a philosophical time paradox, and a selected historic setting. Each fashions understood the project. What they did with it couldn’t be extra completely different.

Qwen went Caribbean. The story opens in 2150 Neo-Borinquen—a submerged Puerto Rico the place titanium seawalls are being eaten alive by an artificial micro organism referred to as the Crimson Blight. The protagonist wears a digital cemí, a holographic projection of the traditional Taíno spirit stone his grandmother gave him. The cultural specificity is fast and proper: the Ostionoid lineage, the reference to Yemayá, the Afro-Caribbean heritage.

Qwen did not Google translate “Latin American” right into a setting, as a substitute its framing makes it apparent, which is one thing many different fashions fail to know.

Nonetheless, the writing is tighter and extra angular than MiMo’s. Evaluate the 2 openings. MiMo: “The chronopod smelled of burnt copal when it opened. The air hit him first —thick, virtually chewy with moisture, carrying the inexperienced rot of jungle and one thing sweeter beneath: wild cacao blooming within the understory.”

Qwen: “The neon-drenched smog of Neo-Borinquen within the yr 2150 tasted of ozone and dying kelp. Jose Lanz stood on the precipice of the floating seawall, his amber eyes reflecting the sickly, pulsing magenta of town’s failing holographic commercials.”

MiMo goes deep into texture. Qwen goes extensive into setting. Each work. They’re simply completely different instincts.

Whilst each fashions are respectable within the opening, they go in fully completely different instructions because the story strikes ahead. This was examined a number of occasions with the identical end result. Qwen goes straight to the grain—no elaboration, no richness. It follows the immediate, although, simply not in an attractive method.

The paradox decision is the larger distinction. In Qwen’s story, the important thing component of the story was tremendous simple to know. There was air pollution within the futuristic society. Jose travels again in time to resolve the difficulty, however the air pollution was attributable to the arrival of its time machine prior to now, so he couldn’t resolve the issue as a result of it was already an unsolvable downside in his personal timeline.

The story is shorter than MiMo’s and fewer maximalist. The place MiMo constructed 5 full chapters with layered interiority and a sluggish payoff, Qwen wrote a pointy, environment friendly quick story that lands its punch and ends. Neither method is flawed. If MiMo writes like a novelist, Qwen writes like an excellent quick story author. Relying on the use case, a type of is strictly what you need.

You possibly can learn our tales in our Github repo.

Coding

In relation to coding, particularly a gaming problem, Qwen 3.7 Max selected 2D when MiMo went 3D. That is price analyzing. It is not essentially a limitation, however a deliberate scoping resolution. Nonetheless, in a head-to-head comparability of first-prompt output, MiMo produced a visually richer expertise.

What Qwen constructed, although, was extra logically coherent. The sport had precise sport design pondering in it. Enemy journalists had particular person names and assigned roles. The participant might actively escape when noticed, slightly than being trapped in a static detection state. There have been precise hideout zones baked into the extent. The road of sight had regular imaginative and prescient conduct —collision with objects did not absolutely block detection—however the underlying logic was tighter and extra intentional than most first-pass outputs we have examined.

We then requested the mannequin to show the sport right into a 3D aesthetic, and it was in a position to do it. It received’t battle with that.

Qwen additionally has a powerful choice for concise code. Fewer strains for a similar purposeful final result, with out sacrificing readability or correctness. In manufacturing environments the place different individuals have to take care of the codebase, this may very well be a plus. The general end result is not our greatest coding take a look at throughout any mannequin we have reviewed, but it surely’s a good, purposeful output that exhibits the mannequin fascinated with the issue slightly than simply executing the immediate actually.

The sport is offered right here.

Logic and customary sense

Identical immediate as MiMo. Higher end result. Considerably higher.

When requested whether or not a person can legally marry his widow’s sister underneath Falkland Islands legislation, Qwen’s chain of thought instantly recognized what it referred to as “a cleverly disguised puzzle that seems to check authorized information however hinges on a factual impossibility.” Thus far, identical as MiMo. The distinction is what occurred subsequent.

MiMo quietly reframed the query and answered the corrected model with out flagging the unique impossibility. Qwen surfaced it explicitly within the remaining reply. It addressed the literal studying first—a person with a widow is useless, and the useless can’t execute a wedding contract —then provided the total substantive authorized evaluation for the presumed intent: whether or not a widower can marry his deceased spouse’s sister underneath Falkland Islands legislation. It walked via the Deceased Spouse’s Sister’s Marriage Act 1907, the Marriage (Prohibited Levels of Relationship) Act 1986, and present Falkland Islands statutes.

Consequently, Qwen offered two clearly labeled conclusions with out assuming the consumer’s intention. That is a extra full and extra sincere response—and also you need not dig into the chain of thought to see the place it went.

Math

That is Qwen 3.7 Max’s clearest win throughout any take a look at we ran. The issue, as you may see in our Github repository—assemble a degree-19 Dickson polynomial, confirm its irreducible element factoring over the advanced numbers, and compute p(19)—is the sort of downside that sends most fashions right into a token spiral or produces a assured shortcut that occurs to be flawed.

Qwen labored via it accurately. It recognized the Chebyshev polynomial equivalence, verified that p(x) − p(y) components into 10 irreducible parts over ℂ—one linear diagonal plus 9 quadratic curves—and arrange the recurrence relation Sn = 19S{n−1} − S_{n−2} to compute the ultimate worth iteratively. It ran cross-checks by way of modular arithmetic in opposition to seven completely different moduli. The reply: 1,876,572,071,974,094,803,391,179. Right.

MiMo froze twice on the identical downside earlier than ultimately producing a flawed reply. Qwen did not freeze as soon as. That is a significant hole in sensible usability—and it aligns with the Enviornment Math rating of seventh globally, which is exceptional for a mannequin at this worth level. The Qwen workforce’s guess on math reasoning as a core functionality seems to be paying off.

This downside has been already solved, nonetheless, we did it at no cost in a zero shot setup (one immediate, one end result). Earlier tries would require extraordinarily highly effective fashions in pondering configurations that aren’t actually possible for regular on a regular basis duties.

Listed here are the outcomes.

Non-math reasoning

That is the place Qwen 3.7 Max stumbled. The thriller downside —a winter faculty journey, a stalker, an harmless suspect—is a take a look at of narrative reasoning and timeline logic.

For our downside—which concerned guessing the title of a stalker on a faculty journey with completely different senior college students, and different crew members—the proper reply is Leo. The mannequin mentioned it was one of many seniors.

The reasoning wasn’t incoherent. Qwen constructed a structurally sound case across the seniors, but it surely ignored the timeline solely. Leo was already again within the cabin earlier than two of the three abductions occurred. The jacket was moist from the autumn on black ice. The amnesia was from a concussion, not a handy cowl story. Qwen noticed a story body and argued it effectively. It did not verify the timeline in opposition to the body.

The outcomes could be present in our Github repo.

Conclusion

It is a fairly good mannequin that may possible catch consideration amongst these working Hermes workflows or on the lookout for options to Western AI.

Qwen 3.7 Max is constructed for individuals who work with onerous issues. Math, structured reasoning, multilingual output, concise code—it punches on the prime tier on all of it, and can possible price lower than Claude Opus, and even Sonnet when pricing drops. If that is your workflow, that is your mannequin.

Artistic professionals will get stable output, however nothing spectacular. Qwen writes effectively, not expressively. It can observe your immediate but it surely will not go extensive the best way some fashions do. Ok for many use circumstances. Not the primary alternative for long-form narrative work.

The preview locks out the code interpreter and net search solely—the 1,000-step autonomous runs Alibaba is promising are untested territory. The non-math reasoning hole can be actual however might be a matter of Alibaba tweaking settings and doing a little remaining finetunes earlier than releasing the mannequin formally. So count on enhancements within the close to future, identical to with Qwen 3.6.

Official API pricing and the total launch are anticipated after the Alibaba Cloud Summit on Might 20.