Alibaba announces Qwen3.8-Max, its most powerful AI model and the first in Max class with open weights

Alibaba has announced Qwen3.8-Max, its most powerful language model to date with 2.4 billion parameters, which will arrive next week as the first Max model with open weights, as well as with great capabilities for programming and collaborative engineering.

The Artificial Intelligence (AI) laboratory behind Qwen3.8-Max has shared all the details of the new model on its website, along with the announcement of the lighter version Qwen3.8-27B, which will also roll out next week.

As the company has explained, among the key points of Qwen3.8-Max stands out its ability to develop complete projects from scratch for more than 10 days without human intervention. It also excels for its ability to plan long-term (for example, a strategy for a project like a year-long online store) and for its native multimodal abilities, which use real-time vision to correct and adapt as it works.

Alibaba itself frames Qwen3.8-Max as comparable to leading frontier systems and only behind Fable 5, although we have to wait for official ‘benchmarks’ to contrast its performance beyond the company’s statements.

In the benchmarks published by Qwen, its new frontier model stands out above all in the multimodal field, with 91.9 in LogicVista and 88.0 in RealWorldQA, and maintains the level in general capabilities and programming with 92.6 in GPQA Diamond and 86.6 in Terminal Bench 2.1

In these ‘benchmarks’, compared to Fable 5 and GPT 5.6 Sol, Qwen3.8-Max ties with Anthropic’s AI model in GPQA Diamond and it is behind OpenAI, which reaches 94.1, while in Terminal Bench 2.1, Fable 5 remains at 84.6 and GPT 5.6 Sol continues to lead with 88.8.

Alibaba has announced the prices of Qwen3.8-Max on its X account (formerly Twitter). Per entrance has a cost of 2 dollars per million tokens and 6 dollars per million output tokens, meanwhile, the implicit cache (reuses repeated text) remains at 0.25 dollars per million tokens.

QWEN3.8-MAX EVOLVES ON ITSELF

In programming tasks, according to the technology company, Qwen3.8-Max sets a new standard, since “it is not limited to following a fixed plan. Evolves on its own through feedback loops, whether building a self-refining testing environment or refining a research method experiment after experiment.”

Likewise, the new frontier model managed to accumulate 265 ‘commits’ (commitments of changes), 127 pull requests and 151 incidents (‘issues’) in approximately 16 days of fully autonomous operation in the ‘oh-my-cli’ coding experiment available on GitHub, demonstrating its continually evolving autonomous programming capabilities.

Upon the release of Qwen3.8-Max open weights, any developeryou will be able to download it, run it on your own servers and adapt it to your needs starting next week. It is the first time that Qwen has done so with a Max model, the most powerful, and which until now was reserved only for its smaller versions.