Alibaba releases Qwen3.8-Max, a 2.4-trillion-parameter model

Background

Qwen is the large language model family developed by the Qwen team at Alibaba Cloud. The family includes open-weight models that anyone can download and run, and proprietary models offered through Alibaba's cloud services. Qwen3.8 was the newest generation in this line in August 2026, and it arrived in two forms: a cloud model called Qwen3.8-Max and a separate set of open weights ^[1].

What happened

Alibaba previewed Qwen3.8-Max, a model with 2.4 trillion parameters, in July 2026. The cloud version was released on 3 August 2026 ^[1]. On 12 August 2026 the company released open weights for Qwen3.8-2.4T-A95B, a checkpoint with 2.4 trillion total parameters, of which 95 billion are activated for each token ^[1][2].

Parameters are the numerical weights a model learns during training. A model's total parameter count describes its capacity, but a mixture-of-experts design means that only a fraction of those parameters do work on any single token. The "A95B" in the checkpoint name refers to that active share.

Key details

The open checkpoint has 92 layers. The layers alternate between two kinds of block, Gated DeltaNet and Gated Attention, and each block is followed by a mixture-of-experts layer with 512 experts. For each token, 10 routed experts and 1 shared expert are active. The checkpoint also uses multi-token prediction, a training technique in which the model learns to predict more than one upcoming token at each position ^[2].

The open checkpoint is text-only, and its thinking mode is always on, meaning the model writes out a reasoning process before answering. Qwen3.8-Max, the cloud model, goes further: it accepts image input, offers a non-thinking mode, uses a default context window of one million tokens, and includes built-in tools ^[2].

The open weights are released under a custom license rather than an open-source license such as Apache 2.0. Wikipedia reports that providers with more than 50 million US dollars in annual revenue need a commercial license from Alibaba to use the model ^[1][2].

Alibaba's reported results for Qwen3.8-Max include 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 92.6 on GPQA Diamond and 43.6 on Humanity's Last Exam (HLE), rising to 56.2 when the model may use tools. Terminal Bench and SWE-bench Pro test coding and software-repair tasks, GPQA Diamond tests graduate-level science questions, and HLE is a set of very hard questions written by experts across many fields. These are figures reported by the company ^[2].

Impact

The release gives developers two routes into the same generation of models. Companies and individuals who need the largest scale can use the cloud version, while those who want to run or study the weights can download the open checkpoint, subject to the custom license. For readers following the field, the key point is scale. A model of this size is far beyond what an ordinary personal computer can run, so most people will reach it through a service or a hosted deployment, even when the weights are public.