Alibaba’s Qwen Team launches Qwen3-Max-Preview with 1 trillion parameters, but unlike past releases, it skips open source—signalling a major strategic shift.
Alibaba’s Qwen Team has unveiled Qwen3-Max-Preview (Instruct), its largest large language model yet, boasting 1 trillion parameters. This marks a departure from the lab’s open source tradition, as the model is only available via Alibaba Cloud API, Qwen Chat, OpenRouter, and Hugging Face’s AnyCoder—not under an open source licence.
Previously, Qwen released open source models that rivalled Western AI labs. The shift to controlled distribution raises questions about the future of open access in trillion-parameter-scale models.
In benchmarks, Qwen3-Max-Preview outperformed the company’s own Qwen3-235B-A22B-2507 and led across SuperGPQA, AIME25, LiveCodeBench v6, Arena-Hard v2, and LiveBench (20241125), ranking ahead of Claude Opus 4, Kimi K2, and Deepseek-V3.1.
Qwen’s official post described it as the lab’s “biggest yet,” adding that “scaling works — and the official release will surprise you even more.” Binyuan Hui, Staff Research Scientist at the Qwen Team, confirmed: “Qwen-Max has successfully scaled to 1T parameters and development is still moving forward.” He hinted at more updates “as soon as next week” and noted that reasoning features are “on the way.”
The model supports a 262,144-token context window, context caching, and is designed for reasoning, coding, structured data handling, creative tasks, and agentic behaviours. Pricing is tiered, starting at $0.861 per million input tokens for smaller workloads.
Early testers reported blazing-fast responses, sometimes faster than ChatGPT, and unexpected reasoning-like behaviour. Hugging Face’s Ahsen Khaliq showcased its creative power by generating a voxel pixel garden in a single prompt.
While the preview demonstrates Alibaba’s commitment to ultra-large models, its closed distribution highlights a strategic pivot—from open innovation to monetisation—with implications for both the developer community and enterprise adoption.
