Alibaba Cloud has launched a new Qwen3-Max AI model, which is the largest and most capable AI model from the company. The preview version of the model is out, and it has already secured a top-three position on the global Text Arena leaderboard, outperforming GPT 5 chat.
Interestingly, the official release is said to improve performance in coding and agent tasks. It is expected to deliver amazing results across multiple benchmarks, including knowledge, reasoning, instruction, following human presence, alignment, and multilingual understanding. The company shared results from the SWE bench, where Qwen3-Max-instructor secured 69.6 and was ranked amongst the world’s top coding models. Apart from that, the Tau2-Bench score of 74.8 surpasses big competitors, like cloud ops and DeepSeek V3.1.
However, it’s important to keep in mind that the Qwen3-Max thinking variant is still under training. Yet when it was augmented with two usage and scale test time commute, it achieved perfect 100-point scores on challenging mathematical benchmarks such as AIME 25 and HMMT.
With over 1 trillion parameters and pretraining on 36 trillion tokens, Qwen3-Max uses an optimized Mixture of Experts (MoE) architecture that ensures stable and efficient training. The model also employs advanced strategies like ChunkFlow, SanityCheck, and EasyCheckpoint, which significantly improve long-context training and reduce downtime from hardware issues.
If you want to use the Qwen3-Max instruct, then it’s available via Qwen Chat or through its API on Alibaba Cloud. Interested users can interact with the model directly on the platform or integrate it into applications using the API.
On that note, I’d like to mention our new product, Geekflare Connect. It lets you connect multiple AI models, like OpenAI, Google, Anthropic, and more, in one place. This helps your team stay organised and can reduce AI costs by up to 65%.
