# Qwen3.8-Max: Alibaba's flagship model, what it measures and what it costs

> Qwen3.8-Max from Alibaba: 2.4T parameters, 1M context, API prices and the benchmarks its lab published, plus when to pick it, as of November 15, 2026.

*Tags: ai, models, alibaba, qwen, open-weights*

---


> [!info] Quick Definition
> **Qwen3.8-Max** is Alibaba's flagship model, introduced on August 3, 2026: a mixture-of-experts (MoE) model with 2.4 trillion total parameters and 95 billion active, a 1M-token context window, built for coding and long-running agent tasks. It is the official version based on the Qwen3.8-2.4T-A95B weights.


All data on this page is as of November 15, 2026 and comes from Alibaba pages. Prices and licenses change.

## What is Qwen3.8-Max and when did it come out?

Alibaba announced it on August 3, 2026 as its largest and most capable model to date. It builds on the Qwen3.5 series and uses an MoE architecture: of the 2.4T parameters, only 95B work on each token. For more on this kind of model, read [[small-vs-large-models|small vs. large models]].

| Fact | Value |
|---|---|
| API ID | `qwen3.8-max` |
| Open weights | `Qwen/Qwen3.8-2.4T-A95B` (Hugging Face) |
| Date | August 3, 2026 |
| Context | 1M tokens by default; in the weights, 262,144 natively and up to 1,010,000 with extension |
| API price | USD 2 input / USD 6 output per million tokens (Singapore and US) |
| Open or closed | Weights published under its own license (see below) |

The same Alibaba Cloud documentation also lists `qwen3.8-flash`, described as fast and cost-effective, and `qwen4-max-preview`, listed under a plan called Token Plan. This page covers only Qwen3.8-Max.

## What is it good for?

Alibaba presents it for coding, real-life work, research and long-horizon tasks. Its announcement says the model ran a software engineering project for 16 days and produced an agent framework that Alibaba published on GitHub. That is a claim from the lab, not independent proof.

Use cases the announcement mentions that are close to your work:

- Rebuilding a frontend project from screenshots (with the hosted version, which accepts vision).
- Turning long documents into a searchable knowledge base.
- Building games and animations from natural language.

It is a [[reasoning-models|reasoning model]]: in the open weights thinking is always on, so every answer spends extra tokens thinking. That raises cost and latency on simple tasks.

## Which benchmarks did it publish and how do you read them?

Alibaba measured these scores with its own harnesses, according to the Hugging Face model card. They are not suitable for ranking against models from other labs. To see why, read [[ai-benchmarks|how to read AI benchmarks]].

| Benchmark | Score | Who measured it and how |
|---|---|---|
| Terminal Bench 2.1 | 86.6 | Alibaba, with Claude Code (average of 10 runs, 5-hour limit) |
| SWE-bench Pro | 67.7 | Alibaba, with the Claude Code harness, on a refined version of the benchmark (problematic tasks corrected) |
| DeepSWE 1.1 | 56.6 | Alibaba (reports the best result across two harnesses) |
| WideSearch | 81.9 | Alibaba, with Qwen-Agent (average F1 over 4 runs) |
| GPQA Diamond | 92.6 | Alibaba |
| IFBench | 82.8 | Alibaba |
| HLE | 43.6 | Alibaba |
| HLE with tools | 56.2 | Alibaba |

Two lessons for a junior:

- **The name misleads.** Alibaba reports Terminal Bench 2.1; Mistral and Moonshot report Terminal-Bench 4.0. Those are different versions: do not compare their numbers.
- **The harness matters.** The card notes the model does best with Claude Code on DeepSWE 1.1. The same model scores differently depending on the program wrapped around it.

The card also includes Alibaba's internal tests (for example QwenReactBench and QwenSVGBench). Because the lab designed and runs them, treat them as a signal of where it put its effort, not as a standard.

## How much does it cost and how much context does it have?

According to the Alibaba Cloud Model Studio pricing documentation (as of November 15, 2026), in dollars per million tokens, the same in thinking and non-thinking mode:

| Region | Input | Output |
|---|---|---|
| Singapore (international) and US (US deployment) | 2 | 6 |
| Hong Kong, Germany, Japan and US (global) | 1.65 | 4.951 |

There is a single tier: up to 1M tokens per request, all at the same price. Input prices do not include caching. Alibaba's GitHub README states 1M context, with a maximum input of 991.80K and a maximum output of 131.07K. If these units are new to you, read [[tokens|tokens]] and [[cost-per-token|cost per token]]; for context, [[context-window|context window]].

## When should you pick it and when not?

Pick it when:

- You need a strong coding model through an API and want to compare cost against others.
- You care about being able to download the weights, for example for audit or self-hosting (check the license first; see [[open-vs-closed-models|open vs. closed models]]).

Avoid it when:

- You need it to read images and plan to use the open weights: they are text-only.
- Your task is simple and high-volume: a reasoning model this size spends more than it needs to.
- Your company offers a model service to third parties and exceeds 50 million dollars of revenue in 12 months: the license requires a separate agreement.

## Example

Illustrative example: a product designer has a design system in a JSON of tokens and wants to generate a React button component with its variants. She takes three tasks she already solved with another model, sends them to `qwen3.8-max` through the API with the tokens as context, and notes, for each, the token cost and whether the code compiles on the first try.

What to try tomorrow: run those three tasks with `reasoning_effort` set to `medium` and to `xhigh`, and compare cost and quality. Decide with your own tasks, not with the lab's table.

## Common mistakes

- Comparing Terminal Bench 2.1's 86.6 with another model's Terminal-Bench 4.0 score: they are different versions.
- Assuming that because the weights are on Hugging Face the license is free. It is its own license with conditions.
- Expecting vision in the open weights.
- Using a price from another region: input varies between 2 and 1.65 per million.
- Taking the lab's internal tests as an independent standard.

## Sources

- [Hugging Face: Qwen3.8-2.4T-A95B model card](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B)
- [Hugging Face: Qwen3.8-2.4T-A95B license](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE)
- [Alibaba Cloud: Qwen3.8-Max announcement](https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-8-max-its-largest-and-most-capable-flagship-model-to-date_603420)
- [Alibaba Cloud Model Studio: model pricing](https://www.alibabacloud.com/help/en/model-studio/model-pricing)
- [Alibaba Cloud Model Studio: model list](https://www.alibabacloud.com/help/en/model-studio/models)
- [Alibaba Cloud on GitHub: Qwen3.8-max](https://github.com/AlibabaCloud-Official/Qwen3.8-max)


---

Source: https://www.fernandoux.com/en/wiki/ai/qwen3-8-max/
