Alibaba releases its Qwen3.8 flagship as open weights

Alibaba’s Qwen team has released the flagship of its Qwen3.8 family as open weights. Alibaba Cloud added the model, qwen3.8-2.4t-a95b, to its Model Studio API on August 12, 2026, ten days after it launched the hosted Qwen3.8-Max there. The model card calls it the first Qwen-Max-class model released with open weights.

What it is

Qwen3.8-2.4T-A95B
Architecture	Sparse mixture of experts: 2.4T total, about 95B active
Context	262,144 tokens natively, extendable to 1,010,000
Input / output	Text only
Thinking	Always on; depth set by reasoning_effort (low, medium, xhigh)
License	Listed on Hugging Face as “qwen3.8-max”

The hosted Qwen3.8-Max, by contrast, has native image understanding and a 1M-token context by default.

Qwen reports 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro and 92.6 on GPQA Diamond for the open model. These are the vendor’s own figures.

What it costs to use hosted

International (Singapore) prices on Alibaba Cloud Model Studio, per 1M tokens:

Model	Input	Output
Qwen3.8-2.4T-A95B	$2.00	$6.00
Qwen3.8-Max	$2.00	$6.00
Qwen3.8-Flash	$0.15	$0.47

Each model has a single price band up to 1M tokens, so long prompts don’t cost more per token. Model Studio gives 1M free tokens per model in the Singapore region, valid for 90 days. All three models also support a context caching discount on input tokens.

Who it is for

The open weights let you run the model on your own servers or with a host of your choice, so your prompts don’t have to go to Alibaba. At 2.4 trillion parameters, though, self-hosting it takes a large GPU cluster; for most teams, the hosted API or a smaller model is the practical choice.

What to do

Read the license on the model card before you use the weights in a product. If you only need the API, compare Qwen3.8-Max with Qwen3.8-Flash on your own tasks: Flash costs less than a tenth as much per output token.