Alibaba Cloud offers a wide variety of LLM adapted to the majority of market use cases, for a very competitive price.
China continues to excel in AI, with a new glance signed Alibaba. The firm has just struck strongly on March 5, 2025 with QWQ-32B, a model of reasoning that rivals with the very publicized Deepseek R1, but for a fraction of its computational resources. Where Deepseek R1 requires 671 billion parameters and more than 1,500 GB of VRAM (16 GPU NVIDIA A100), QWQ-32B reaches comparable performance with only 32 billion parameters and 24 GB of true on a single GPU. This engineering feat is part of a wider Alibaba Cloud strategy, which has developed a remarkably complete LLM family since April 2023: Qwen. Its general models, owners as open source, regularly rank among the most efficient in benchmarks, confirming the rise of Chinese AI in the face of American heavy goods vehicles.
Owner models and open source versions
In March 2025, Alibaba Cloud offered both proprietary and open source models. Modality, context, performance on complex tasks, speed … Here is a comparison of all the large Language Model currently deployed on Alibaba Cloud Model Studio, the Alibaba Cloud platform dedicated to generative AI.
Owner models
| Model | Complex tasks | Modality (input) | Tokens (input) | Tokens (output) | Context | Fast |
|---|---|---|---|---|---|---|
| Qwen-max (2.5 max) | X | text | 30 720 | 8 192 | 32,768 | |
| Qwen-Plus | X | text | 129 024 | 8 192 | 131 072 | X |
| Qwen-turbo | text | 1,000,000 | 8 192 | 1,000,000 | X | |
| Qwen-Vl-Plus | text, image | 6,000 | 1,500 | 7,500 | ||
| Qwen-Vl-Max | X | text, image | 6,000 | 1,500 | 7,500 |
Open Source models
| Model | Complex tasks | Modality (input) | Tokens (input) | Tokens (output) | Context | Fast |
|---|---|---|---|---|---|---|
| Qwen2.5-14b-instruct-1m | text | 1,000,000 | 8 192 | 1,000,000 | ||
| Qwen2.5-7b-instruct-1m | text | 1,000,000 | 8 192 | 1,000,000 | X | |
| Qwen2.5-72b-instruct | X | text | 129 024 | 8 192 | 131 072 | |
| Qwen2.5-32b-instruct | X | text | 129 024 | 8 192 | 131 072 | |
| Qwen2.5-14b-instruct | text | 129 024 | 8 192 | 131 072 | ||
| Qwen2.5-7b-instruct | text | 129 024 | 8 192 | 131 072 | X | |
| Qwen2-72b-instruct | X | text | 128,000 | 6 144 | 131 072 | |
| QWEN2-57B-A14B-Instruct | text | 63,488 | 6 144 | 65 536 | ||
| Qwen2-7b-instruct | text | 128,000 | 6 144 | 131 072 | X | |
| Qwen1.5-110b-chat | text | 6,000 | 2,000 | 8,000 | ||
| Qwen1.5-72B-Chat | text | 6,000 | 2,000 | 8,000 | ||
| Qwen1.5-32B-Chat | text | 6,000 | 2,000 | 8,000 | ||
| Qwen1.5-14b-chat | text | 6,000 | 2,000 | 8,000 | ||
| Qwen1.5-7B-Chat | text | 6,000 | 2,000 | 8,000 | X | |
| Qwen2.5-VL-72B-Instruct | X | text, image, video | 129 024 | 8 192 | 131 072 | |
| Qwen2.5-VL-7B-Instruct | text, image, video | 129 024 | 8 192 | 131 072 | X | |
| Qwen2.5-VL-3B-Instruct | text, image, video | 129 024 | 8 192 | 131 072 | X | |
| QWQ-32B | X | text | Nc | Nc | 131 072 |
The open source versions of Qwen constitute an interesting alternative for companies concerned with their technological independence or confronted with strict regulatory constraints (in local use therefore). For demanding tasks, Qwen2.5-72B-Instruct stands out with its 72 billion parameters, while QWEN2.5-7B-instruct effectively meets textual current needs with higher speed. Multimodal models like QWEN2.5-VL-72B-Instruct add the ability to process images and videos for up to 10 minutes. Qwen1.5 is, in our view, more relevant for production in production.
Finally, for companies wishing to develop agent capacities, QWQ-32B is the ideal solution. Unveiled on March 5, 2025, the model reasoned as well as Deepseek R1 with only 32 billion parameters.
Very attractive pricing
| Model | Price of 1000 tokens in Input ($) | Price of 1000 tokens in output ($) |
|---|---|---|
| Qwen-max (2.5 max) | 0.0016 | 0.0064 |
| Qwen-Plus | 0.0004 | 0.0012 |
| Qwen-turbo | 0.00005 | 0.0002 |
| Qwen-Vl-Plus | 0.00021 | 0.00063 |
| Qwen-Vl-Max | 0.0008 | 0.0032 |
| Qwen2.5-14b-instruct-1m | Free (for the moment) | Free (for the moment) |
| Qwen2.5-7b-instruct-1m | Free (for the moment) | Free (for the moment) |
| Qwen2.5-72b-instruct | Free (for the moment) | Free (for the moment) |
| Qwen2.5-32b-instruct | Free (for the moment) | Free (for the moment) |
| Qwen2.5-14b-instruct | Free (for the moment) | Free (for the moment) |
| Qwen2.5-7b-instruct | Free (for the moment) | Free (for the moment) |
| Qwen2-72b-instruct | Free (for the moment) | Free (for the moment) |
| QWEN2-57B-A14B-Instruct | Free (for the moment) | Free (for the moment) |
| Qwen2-7b-instruct | Free (for the moment) | Free (for the moment) |
| Qwen1.5-110b-chat | Free (for the moment) | Free (for the moment) |
| Qwen1.5-72B-Chat | Free (for the moment) | Free (for the moment) |
| Qwen1.5-32B-Chat | Free (for the moment) | Free (for the moment) |
| Qwen1.5-14b-chat | Free (for the moment) | Free (for the moment) |
| Qwen1.5-7B-Chat | Free (for the moment) | Free (for the moment) |
| Qwen2.5-VL-72B-Instruct | Free (for the moment) | Free (for the moment) |
| Qwen2.5-VL-7B-Instruct | Free (for the moment) | Free (for the moment) |
| Qwen2.5-VL-3B-Instruct | Free (for the moment) | Free (for the moment) |
| QWQ-32B | Nc | Nc |
Qwen stands out as an alternative solution to American models. Combining technical performance, diversity of supply (proprietary models and open source) and prices up to three times lower than the competition, Qwen perfectly illustrates the rise of China in the field of artificial intelligence. A suitable value proposal adapted to the majority of use cases in business. Alibaba Cloud is now a key player in the world generative AI market.