Alibaba Launches Trillion-Parameter AI Model Qwen-3-Max-Preview

Alibaba Group Holding has released Qwen-3-Max-Preview, its largest artificial intelligence model to date and the company’s first to exceed one trillion parameters. The new model places Alibaba among major players like OpenAI and Google DeepMind in the development of trillion-parameter models.

Credit: QWEN.AI

Qwen-3-Max-Preview was released on Alibaba’s official cloud-services platform and the large language model marketplace OpenRouter on Friday. It is the latest addition to the Hangzhou-based company’s Qwen3 series, which was first released in May.

The Qwen3 series initially included models ranging from 600 million to 235 billion parameters. Parameters are variables that encode an AI system’s intelligence and are adjusted during the training process.

Generally, a higher number of parameters indicates stronger capabilities, though it requires more computational power to train and run the model. OpenAI’s GPT-4.5, an industry-leading model, has an estimated parameter count of five to seven trillion.

Alibaba stated Qwen-3-Max-Preview is a text-only model that outperformed its previous "best" model, the Qwen3-235B-A22B-2507, in internal tests. The Qwen3-235B-A22B-2507 was released in July.

The company also published scores suggesting that Qwen-3-Max-Preview surpassed MoonShot AI’s Kimi K2, a non-reasoning version of Anthropic’s Claude Opus 4, and DeepSeek V3.1 on five benchmarks. These scores were not part of an official technical report for the model.

Alibaba stated, "Qwen3-Max-Preview shows substantial gains … in overall capability, with significant enhancements in Chinese-English text understanding, complex instruction following, handling of subjective open-ended tasks, multilingual ability, and tool invocation." The company added, "Scaling works – and the official release will surprise you even more."

Alibaba’s Qwen models have propelled the company to the top position in the global open-source AI ecosystem. They have garnered more than 20 million downloads and 100,000 derivative models on open-source developer platform Hugging Face.

However, Qwen-3-Max-Preview has not been open-sourced, with access currently available only through official channels. Qwen2.5-Max, the company’s previous model in the Max series released in January, was also not open-sourced.

Alibaba artificial intelligence engineer Binyuan Hui indicated on social media that a "thinking" version of the model is "on the way."

On Alibaba Cloud, Qwen-3-Max-Preview has tiered pricing, starting at USD 0.861 per million input tokens, and USD 3.441 per million output tokens. This makes it one of the most expensive Qwen models for access via an application programming interface.

For comparison, the non-thinking version of Qwen3-235B-A22B-2507 charges USD 0.287 per million input tokens, and USD 1.147 per million output tokens.

Credit: KIMI

Kimi K2’s new model charges USD 0.60 per million input tokens, and USD 2.50 per million output tokens.

Alibaba has committed 380 billion yuan (USD 52 billion) to artificial intelligence infrastructure investments over the next three years. This commitment exceeds its investment in this sector over the past decade.

The company is already observing returns on its artificial intelligence investment. Artificial intelligence-related products have achieved triple-digit growth for the past eight consecutive quarters, according to its latest financial results.

---

• Alibaba released Qwen-3-Max-Preview, its largest artificial intelligence model, featuring over one trillion parameters.

• The model, part of the Qwen3 series, was launched on Alibaba’s cloud platform and OpenRouter.

• Qwen-3-Max-Preview outperformed previous Alibaba models and rivals like Kimi K2 in internal tests.

Source: SCMP