Alibaba Launches Qwen2.5-Max, Claims It Beats DeepSeek and GPT-4o

Alibaba has unveiled Qwen2.5-Max, a large language model that, according to the company's own testing, outperforms DeepSeek-V3 and GPT-4o on several benchmarks. The announcement, published by the Qwen team, comes at a pointed moment: just days after DeepSeek upended the global AI conversation, and right in the middle of Chinese New Year celebrations. The underlying message is clear — the race for the AI frontier is no longer a single Chinese lab's game, but an entire ecosystem pushing forward at once.

What Qwen2.5-Max is

It's a Mixture-of-Experts (MoE) model, an architecture that, instead of activating the entire network for every query, splits the work among specialized "experts" and only switches on the ones needed. This lets developers scale up model size without blowing up compute costs on every response.

According to Alibaba, Qwen2.5-Max was pretrained on more than 20 trillion tokens — the smallest unit of text a model processes, roughly equivalent to word fragments — and then refined using two standard post-training techniques: supervised fine-tuning (SFT), which trains the model on curated examples of correct answers, and reinforcement learning from human feedback (RLHF), which adjusts its behavior based on human evaluators' preferences.

The company itself acknowledges in its announcement just how much DeepSeek has shaped the landscape. The Qwen team admits that the research and industry community had "limited experience" effectively scaling extremely large models, and that many critical details of that process only came to light with the recent release of DeepSeek V3. It's an unusual move to cite a direct rival as a technical reference point.

What the benchmarks show

Alibaba compares Qwen2.5-Max on two fronts: chat-tuned models (the ones that reach end users) and base models (the pre-polish version).

In the instruct-model comparison — the one that matters for applications like chat or coding — Alibaba pits Qwen2.5-Max against DeepSeek V3, GPT-4o, and Claude-3.5-Sonnet. According to its results, the model outperforms DeepSeek V3 on tests including:

• Arena-Hard, which approximates human preferences.
• LiveBench, which measures general capabilities.
• LiveCodeBench, focused on coding.
• GPQA-Diamond, made up of advanced-level questions.

The company adds that it gets "competitive" results on other evaluations such as MMLU-Pro, which tests knowledge through college-level problems. The distinction matters: "competitive" is not the same as "superior," and the announcement itself is careful to separate the benchmarks where it claims a win from those where it merely comes close.

On base models, Alibaba acknowledges it cannot access proprietary systems like GPT-4o and Claude-3.5-Sonnet, so the comparison is limited to open-weight models: DeepSeek V3, Llama-3.1-405B — the largest open-weight dense model — and Qwen2.5-72B, one of the top open-weight dense models. There, the company claims "significant advantages across most benchmarks."

As always, these figures deserve some caution: they're internal evaluations from the model's own maker, published without an independent technical report to back them up at the time of the announcement. A lab declaring victory on its own tests is the industry norm, not the exception, and it's no substitute for outside validation.

How to use it, and what the entry price means

Qwen2.5-Max is already available on Qwen Chat, where users can talk to the model directly and try features like artifacts and search. For developers, the API is offered through Alibaba Cloud under the model name qwen-max-2025-01-25: it requires registering an Alibaba Cloud account, activating the Model Studio service, and generating an API key.

One detail that matters for anyone looking to integrate it: Qwen's APIs are compatible with OpenAI's. In practice, that means a company already running its code on OpenAI's tools can point it at Qwen2.5-Max by changing little more than the server address and the model name. That compatibility lowers the cost of switching providers and is an obvious commercial lever in a market where keeping developers locked in is half the battle.

Also worth noting is the example Alibaba chose to showcase the model in its sample code: asking which number is larger, 9.11 or 9.8. It's exactly the kind of simple numerical comparison that has tripped up several language models before, and using it as a demo is a nod to that well-known weak spot.

Why this announcement matters

The context tells most of the story. For the past several years, the dominant narrative placed the frontier of large models in the hands of a handful of labs, mostly American. DeepSeek's emergence broke that frame by putting a Chinese player squarely in the front-line conversation. Qwen2.5-Max is Alibaba's answer: not a lone lab, but a major tech company with its own cloud pushing into the same territory — and doing so just days later, right in the middle of Chinese New Year.

The Qwen team makes its underlying ambition clear. It says it's committed to improving its models' reasoning capabilities through reinforcement learning at scale, and claims this effort "holds the promise of enabling our models to transcend human intelligence." It's a grandiose statement, the kind that's common in industry announcements, and one worth separating from what the model actually demonstrates today.

For users and businesses, the tangible takeaway is more modest but more useful: there's a new large-scale model accessible via chat and API, with technical compatibility that makes it easy to try, and with the promise — so far backed only by the company's own benchmarks — of performing at the level of the best proprietary systems. What remains to be seen is how it behaves outside the controlled terrain of benchmark testing, and whether future independent evaluations confirm the numbers Alibaba is putting on the table.