# What is Qwen2.5-Max, and how does Alibaba's model compare with DeepSeek-V3?

Date: 2025-01-30 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-qwen-2-5-max/

**Answer:** Qwen2.5-Max is a large mixture-of-experts language model from Alibaba's Qwen team, released at the start of the Lunar New Year (January 28–29, 2025), trained on more than 20 trillion tokens and offered through the Qwen Chat website and Alibaba Cloud's API. On the five tests Alibaba published, it edges out DeepSeek-V3 and beats GPT-4o on most, while Claude 3.5 Sonnet still leads on three of them.

Alibaba, the Chinese e-commerce and cloud company, has released **Qwen2.5-Max**, the most powerful model so far in its Qwen family. The Qwen team posted the announcement late on January 28, Beijing time, and Reuters reports the release on Wednesday, January 29. That was the opening day of the Lunar New Year, a holiday most people in China spend at home with family.

That timing is telling. Over the past three weeks, the startup DeepSeek has gone from little-known lab to global headline, with a chat assistant and models whose low costs startled Silicon Valley and pulled tech stocks down. Reuters describes Alibaba's holiday release as a sign of the pressure DeepSeek is putting on its Chinese rivals, not only on American ones. Alibaba's message is direct: its cloud unit says the new model outperforms GPT-4o, DeepSeek-V3 and Meta's Llama 3.1 405B almost across the board.

## What Alibaba released

Qwen2.5-Max is a **mixture-of-experts** (MoE) model. Instead of one giant network that works on every word, an MoE model contains many smaller expert networks and a router that sends each piece of text to a few of them. That lets a model hold a great deal of knowledge while keeping each answer affordable to compute. DeepSeek-V3 uses the same general design.

According to the Qwen team:

- It was pre-trained on **more than 20 trillion tokens** (a token is a chunk of text, often part of a word).
- It was then refined with **supervised fine-tuning**, learning from carefully chosen example answers, and **reinforcement learning from human feedback**, where people's preferences between answers guide further training.
- It is available in **Qwen Chat**, Alibaba's chatbot website, and through Alibaba Cloud's API under the model name `qwen-max-2025-01-25`.

Two things are missing from the announcement. Alibaba does not give the model's size in parameters, and it does not offer the weights for download. Unlike DeepSeek-V3, which anyone can download, this one is available only as a service.

The Qwen team is also candid about why it is publishing now. It writes that the industry has had little experience scaling extremely large models, and credits DeepSeek's V3 report with recently making many of those details public.

## How it compares

Alibaba published scores for five tests, comparing the chat-ready version of Qwen2.5-Max with four rivals. Here is what each test measures and the numbers from Alibaba's chart:

| Test | What it measures | Qwen2.5-Max | DeepSeek-V3 | GPT-4o | Claude 3.5 Sonnet |
|---|---|---|---|---|---|
| Arena-Hard | How often answers are preferred, approximating human judges | 89.4 | 85.5 | 77.9 | 85.2 |
| MMLU-Pro | College-level knowledge questions | 76.1 | 75.9 | 77.0 | 78.0 |
| GPQA-Diamond | Hard graduate-level science questions | 60.1 | 59.1 | 53.6 | 65.0 |
| LiveCodeBench | Recent coding problems | 38.7 | 37.6 | 35.1 | 38.9 |
| LiveBench | A broad mix of general abilities | 62.2 | 60.5 | 56.0 | 60.3 |

Read the table carefully and a more nuanced picture appears than the headline. Qwen2.5-Max is ahead of DeepSeek-V3 on all five tests, but often by about a point. It beats GPT-4o on four of five. Claude 3.5 Sonnet, Anthropic's model, scores higher on MMLU-Pro, GPQA-Diamond and LiveCodeBench. These are also Alibaba's own measurements, on tests Alibaba chose, so independent results will matter.

The team also compared its **base model**, the raw version before chat training, with open models such as DeepSeek-V3 and Llama 3.1 405B, and reports an advantage on most tests. It could not compare base models with GPT-4o or Claude, because those companies do not make their base models available.

> **Try it:** Pick one question you know the answer to and that needs a little care, such as "Which is larger, 9.11 or 9.8?", which Alibaba itself uses as its sample prompt. Ask it to two different chatbots and compare not just the answer but the explanation. A small personal test like this tells you more about a model's habits than a one-point benchmark gap.

## Why it matters

Qwen2.5-Max shows how quickly competition is moving among Chinese AI companies. Reuters notes that DeepSeek's earlier, very cheap V2 model set off a price war in China last year, with Alibaba's cloud unit cutting prices by up to 97%, and that ByteDance updated its own flagship model just days after DeepSeek's R1. A major release on a national holiday suggests no one wants to fall behind for even a week.

For builders, it adds one more strong option that speaks a familiar language: Alibaba's API follows the same format as OpenAI's, so existing code can switch with small changes. For everyone else, the lesson is that "best model" claims now change week to week, and each one deserves a careful look at the numbers behind it.

## How can you try Qwen2.5-Max?

The easiest way is Qwen Chat, Alibaba's chatbot website, where Qwen2.5-Max is now available. Alibaba says you can chat with it directly, use search and create small interactive pieces it calls artifacts. As with any online assistant, avoid entering private or sensitive information.

If you write code, register for an Alibaba Cloud account, activate the Model Studio service and create an API key in the console. Because the API is compatible with OpenAI's, you can use OpenAI's Python library and simply point it at Alibaba's address. A first request might look like this:

```prompt
System: You are a helpful assistant.
User: Which number is larger, 9.11 or 9.8? Explain your reasoning in two sentences.
```

Compare the reply with what DeepSeek or ChatGPT says to the same prompt. If all three agree and explain clearly, you have learned something useful about all of them.

## Sources

- Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model: https://qwenlm.github.io/blog/qwen2.5-max/
- Alibaba releases AI model it says surpasses DeepSeek: https://www.reuters.com/technology/artificial-intelligence/alibaba-releases-ai-model-it-claims-surpasses-deepseek-v3-2025-01-29/