CVCOVANewsComputing, Opportunities, Value & Analysis
Artificial Intelligence4 min read

Alibaba launches Qwen4-Max: a giant, cheaper model, still without independent testing to back it up

The Chinese company unveiled a frontier artificial intelligence model with lower prices and compatibility with OpenAI and Anthropic tools. For developers in Latin America it is another option, though questions remain about real-world performance, privacy and access from the region.

By COVA News newsroom
turned-on MacBook Pro with programming code on display

Photo: Arnold Francisca on Unsplash

A developer in Bogotá, Lima or São Paulo who already works with OpenAI or Anthropic tools could, with a configuration tweak, start using one of the largest artificial intelligence models announced to date. That is the proposition of Qwen4-Max, Alibaba's new model, whose general release was dated November 16, 2026, according to the company's changelog cited by CASRAI and the OpenRouter platform. It arrives with eye-catching figures and a lower price than its predecessor. What has yet to arrive are independent evaluations confirming what it promises.

The news topped Techmeme and sparked a wide-ranging debate on Hacker News, the go-to forum for many programmers. The attention has to do with the model's scale. According to MarkTechPost, The New Stack and Quartz, which cites Bloomberg, Qwen4-Max has 2.4 trillion parameters. Parameters are the internal "settings" the system learns during training. It also uses an architecture called mixture of experts: instead of activating the entire model for every word, it switches on only a specialized portion, about 71 billion parameters per chunk of text. This lets an enormous model respond at a considerably lower computational cost.

The same outlets credit it with native vision, meaning the ability to interpret images, and a working memory, or context window, of one million tokens. A token is a piece of a word, and that window is enough to hold several books at once. DataCamp specifies that the maximum input is 991,000 tokens. One caveat is in order: our newsroom's research did not find the 2.4 trillion figure in the providers' listings or in the official thread reviewed, and the official Qwen blog could not be accessed.

Cheaper, according to Alibaba itself

The company's strongest argument is price. According to OpenRouter and several outlets, using the model costs US$1.50 per million tokens sent to it and US$4.50 per million it generates. Qwen3.8-Max, launched in August, cost US$2 and US$6. Documentation for Model Studio, Alibaba's service, shows somewhat lower figures in an unclear table: US$1.413 and US$4.24.

Alibaba also claims that the average cost per task fell 62%, from US$4.43 to US$1.67. That figure comes from Work360, a benchmark designed by the company itself, and compares the new model with its previous version, not with the competition, as Digital Trends pointed out. On that same benchmark, Qwen4-Max scored 52.69 points versus 47.74 for its predecessor, according to Quartz. Other outlets report different savings: Crypto Briefing mentions roughly 50% fewer tokens, and A2Agent around 35% on reasoning-heavy tasks. They measure different things, and all of them come from Alibaba.

The comparison with rivals is even less clear. The company claims its model outperforms Claude Opus 5.5 at programming and competes with Kimi K3.1 Pro. Crypto Briefing gives it 75.7 points on DeepSWE v1.1, a benchmark of software tasks. Quartz, however, identifies the Anthropic rival as "Fable 5.1." The only external reference available points in a different direction: Trending Topics, using data from the evaluation firm Artificial Analysis, places Qwen4-Max ninth, behind Claude, GPT and Kimi models. The commercial blog MindStudio, which tested the preview version on some real tasks, noted that the model rambles less than Qwen 3 and is better at correcting its own mistakes, though without a systematic comparison.

What changes, and what doesn't, for Latin America

There is a shift in strategy that matters. Qwen3.8-Max came with the promise of publishing its weights, the model's downloadable "brain," which allows it to be installed on one's own servers or adapted. Qwen4-Max is proprietary and can only be used over the internet, according to Bloomberg, MarkTechPost and The New Stack. Alibaba says it will later release smaller Qwen4 versions with open weights, though it gave no date.

Access is through QwenCloud and other company tools, or through OpenRouter, where Alibaba Cloud is listed as the only provider. New accounts in Singapore get a free quota of one million tokens for about 90 days, and personal plans start at US$6 a month.

For the region, the picture has gaps. The documentation lists data centers in Beijing, Singapore, Hong Kong, Frankfurt, Tokyo and Virginia, none in Latin America. Nor is it known in which of them this model is available. There is no information on performance in Spanish or Portuguese, on where data is stored, or on compliance with laws such as Brazil's LGPD or Mexican regulations. Local-currency payment options are also unknown.

Qwen4-Max adds another high-end alternative in a race dominated by U.S. companies, and that could push prices down to the benefit of developers in the region. Before moving serious projects, it is advisable to wait for full independent evaluations and privacy terms, and to see whether Alibaba delivers on its promise of open models.

Related news

Artificial Intelligence

Mathematician says he has proved the Riemann hypothesis with help from Claude; it still needs to be verified

Levent Alpöge has published a purported proof of one of the most famous problems in mathematics, produced together with Anthropic's AI. Early reviews are encouraging, but the community has not yet validated it.

Artificial Intelligence

OpenAI cuts GPT-6.2 Astra API pricing by 20%: what changes for those paying in dollars

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol. The flagship model's usage-based pricing drops, but the cheaper model's stays the same, and the fine print will determine how much users in Latin America actually save.

Artificial Intelligence

Nvidia reports earnings without Jensen Huang: what's at stake for AI chips

Five days after its founder's resignation, the company that dominates artificial intelligence chips faces Wall Street. For Latin America, open questions outweigh measurable effects for now.