Alibaba’s Largest Ever AI Qwen 3.8 Max Unveiled With 2.4 Trillion Parameters
Alibaba unveiled Qwen3.8-Max on Monday, describing it as the largest and most capable artificial intelligence model in its Qwen family.
The model has 2.4 trillion total parameters but activates only 95 billion for each request. Alibaba has made it available through Alibaba Cloud’s Model Studio and plans to release its open weights next week, marking the first open-weight release of a Qwen Max-class model.
Close to Kimi K3 in Size
Qwen3.8-Max is close in size to Moonshot AI’s Kimi K3, which launched last month with 2.8 trillion parameters. Kimi K3 is known as the most capable open model so far.
Both models can process text, images and video and support context windows of up to one million tokens. This allows them to handle large amounts of information, including long legal documents, extensive software codebases and hundreds of pages of text.
Qwen3.8-Max uses a mixture-of-experts architecture, which divides work among specialised parts of the model rather than activating all 2.4 trillion parameters for every request. Alibaba said activating only 95 billion parameters helps reduce processing costs and response times.
OpenAI, Anthropic and Google do not publicly disclose the parameter counts of their leading closed-source models.
Low Price
Qwen3.8-Max is also positioned as a relatively inexpensive frontier model. Qwen Cloud lists it at $2 per million input tokens and $6 per million output tokens, with cached input costing considerably less. That puts its API pricing well below many leading proprietary models while offering a context window of around one million tokens.
The model is also available through Qwen’s Token Plan starting at $6 per month, with the current preview promotion providing substantially higher usage capacity.
Highest Rank on Arena.AI
Qwen3.8-Max became the highest-ranked Chinese model for text tasks on the crowdsourced Arena.AI platform shortly after its launch.
It ranked fifth globally with a preliminary score of 1,496, behind Anthropic’s Claude Fable 5 and several Claude Opus variants.
For vision tasks, Qwen3.8-Max ranked second globally with a preliminary score of 1,305. It trailed only Claude Fable 5, which recorded a score of 1,318.
Long-Term Autonomous Work
Alibaba said Qwen3.8-Max can complete complex projects over several days with limited or no human involvement.
In one demonstration, the model spent 16 days developing a self-improving software harness from scratch. It incorporated feedback, ran tests, and repeatedly reviewed its code, previews, and logs during the project.
In another test, Qwen3.8-Max attempted to reproduce a machine-learning research paper from the beginning. Over about 125 hours, it completed 33 rounds of GPU training and wrote around 7,600 lines of code.
The model then proposed 18 improvements that Alibaba said produced better results than the method described in the original paper. When entered into an online competition involving 526 human teams, its result surpassed 87 percent of the field. These results come from Alibaba’s own demonstrations.
Alibaba also showed the model completing a legal compliance review in less than an hour, producing an eight-screen interactive prototype for a digital banking application without revisions, and creating a 26-dish restaurant menu with calorie and cost calculations in a single attempt.
Strong Multimodal and Agent Capabilities
Qwen3.8-Max is designed for more than conventional text generation. Qwen says the model supports reasoning, visual understanding, and text generation, allowing it to work across text, images, and video. It also supports agent-oriented tasks, including tool use and complex coding workflows.
The model is available through Qwen Cloud, where developers can use capabilities such as function calling, structured outputs, web search, code execution and context caching. The latter can reduce both latency and costs when working repeatedly with the same long context.
The post Alibaba’s Largest Ever AI Qwen 3.8 Max Unveiled With 2.4 Trillion Parameters appeared first on ProPakistani.
