Qwen3.8-Max Explained: 2.4T Parameters, Price and Open Weights
In short: Qwen3.8-Max is Alibaba's largest AI model, launched on 3 August 2026: a 2.4-trillion-parameter mixture-of-experts model that uses 95 billion parameters per token. There are two forms. The hosted Qwen3.8-Max on Qwen Cloud costs $2 per million input tokens and $6 per million output, reads text, images and video, and has a 1M-token context window. The open weights, Qwen3.8-2.4T-A95B on Hugging Face, are text-only, always reason before answering, and have a 262,144-token native context extensible to about 1,010,000.
What is Qwen3.8-Max?
Qwen3.8-Max is the flagship of Alibaba's Qwen3.8 model family, aimed at coding, professional work, research and long multi-step agent tasks. Alibaba's Qwen team built it on the architecture of Qwen3.5. Its model card says that "for the first time, Qwen3.8 brings a Qwen-Max-class model to open release" - earlier Max models were API-only.
For what parameters and mixture of experts are, see how LLMs work.
What is the difference between Qwen3.8-Max and Qwen3.8-2.4T-A95B?
They are the same underlying model offered two ways: a hosted API version with extra features, and downloadable weights. The model card describes Qwen3.8-Max as "the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools".
| Qwen3.8-Max (hosted) | Qwen3.8-2.4T-A95B (open weights) | |
|---|---|---|
| Where | Qwen Cloud API, Qwen Studio | Hugging Face download |
| Input | Text, images, video | Text only |
| Thinking | Thinking and non-thinking | Thinking required, cannot be disabled |
| Context | 1M tokens | 262,144 native, extensible to about 1,010,000 |
| Built-in tools | Web search, code interpreter, web extractor, PDF parsing and others | None built in |
| Licence | Proprietary service | Licence named "qwen3.8-max" |
Much launch coverage described "Qwen3.8" as either fully multimodal or text-only without saying which version. Both are true of different versions: the hosted API takes images and video, and the open weights do not.
How much does Qwen3.8-Max cost?
On Qwen Cloud, input costs $2 per million tokens and output $6. From the Qwen3.8-Max page:
| Per 1M tokens | Price |
|---|---|
| Input | $2 |
| Input (implicit cache) | $0.25 |
| Explicit cache creation | $2.50 |
| Explicit cache read | $0.17 |
| Output | $6 |
The open weights are free to download; running a model of 2.4 trillion parameters requires a large multi-GPU server.
What are Qwen3.8-Max's specs?
| Spec | Qwen3.8-Max |
|---|---|
| Developer | Alibaba (Qwen team) |
| Launched | 3 August 2026 (open weights on Hugging Face in August 2026) |
| Parameters | 2.4T total, 95B active per token |
| Architecture | Mixture of experts: 512 experts, 10 routed + 1 shared active; 92 layers |
| Context (hosted) | 1M tokens; max input 991K, max output 131K |
| Maximum reasoning (hosted) | 262K tokens |
| Input (hosted) | Text, images, video |
| Features (hosted) | Function calling, structured outputs, context cache, batches, web search, fine-tuning |
| Rate limits (hosted) | 2M tokens and 15K requests per minute |
| Reasoning control | reasoning_effort; preserve_thinking keeps earlier reasoning in context |
Is Qwen3.8 open source?
The weights are published, under a custom licence named "qwen3.8-max" - not Apache 2.0. The Qwen3.8-2.4T-A95B repository and an FP8 checkpoint are on Hugging Face. A smaller sibling, Qwen3.8-27B, is released under Apache 2.0 and accepts images and video; that licence does not cover the 2.4T weights. The weights work with inference software such as vLLM and SGLang - see how to run AI models locally.
How does Qwen3.8-Max rank?
On LMArena's text leaderboard Qwen3.8-Max has 1482 Elo (rank 24, max effort). Qwen3.7 Max has 1474 Elo (rank 42). The full ranked table is on our benchmarks page.
Words to know
- Mixture of experts (MoE) - a design that uses only some of the model's "experts" for each token.
- Active parameters - the parameters actually used for each token.
- Open weights - model files published for anyone to download and run.
- Thinking mode - the model reasons before answering.
- Native context - the context length a model was built for, before extension techniques.
- Implicit / explicit cache - automatic versus requested reuse of repeated input, billed at lower rates.
Related guides
Frequently asked questions
How many parameters does Qwen3.8-Max have?+
Qwen3.8-Max has 2.4 trillion parameters in total, of which 95 billion are active for each token. It is a mixture-of-experts model with 512 experts, 11 of which are used per token.
How much does Qwen3.8-Max cost?+
On Qwen Cloud it costs $2 per million input tokens and $6 per million output tokens, with implicit-cache input at $0.25 per million. The open weights are free to download.
Is Qwen3.8-Max multimodal?+
The hosted Qwen3.8-Max on Qwen Cloud accepts text, images and video. The open-weight Qwen3.8-2.4T-A95B is text-only; its model card states that multimodal inputs are not supported.
What is Qwen3.8-Max's context window?+
The hosted Qwen3.8-Max has a 1M-token context window with up to 131K tokens of output. The open weights have a 262,144-token native context, extensible up to about 1,010,000 tokens.
Is Qwen3.8 Apache 2.0?+
The 2.4T weights use a custom licence named "qwen3.8-max", not Apache 2.0. The smaller Qwen3.8-27B is released under Apache 2.0.
Can you turn off thinking in Qwen3.8?+
In the open-weight Qwen3.8-2.4T-A95B, no - thinking is required for all interactions. The hosted Qwen3.8-Max supports non-thinking mode.
Enjoyed this? Get more in your inbox.
Up to ten of the week's AI news stories, by email once a week.