# Qwen3.8 2.4T A95B

The open-weight release of Alibaba Cloud's Qwen 3.8 flagship. Qwen3.8 2.4T A95B activates 95B of its 2.4 trillion parameters per token, and its mostly-linear attention stack is what keeps a window of 262.1K tokens affordable to serve.

- **Model ID:** `alibaba/qwen3.8-2.4t-a95b`
- **Type:** chat
- **Providers:** deepinfra, fireworks, gmicloud, novita, modal, runinfra, togetherai
- **Context window:** 262,144
- **Maximum output tokens:** 131,072
- **Pricing:** $1.65/1M input tokens, $4.951/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/qwen3.8-2.4t-a95b

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'alibaba/qwen3.8-2.4t-a95b',
  prompt: 'Why is the sky blue?'
})
```

## About

Qwen3.8 2.4T A95B was released August 3, 2026 as the open-weight version of Alibaba Cloud's Qwen 3.8 flagship, following the hosted Qwen3.8 Max endpoint that opened earlier the same month.

The architecture is a mixture-of-experts design with 2.4 trillion total parameters and 95 billion active per forward pass, spread across 512 routed experts and 92 layers. What distinguishes it is the attention stack: 69 of those 92 layers use linear attention, with full attention layers interleaved at intervals. In the linear layers a growing KV cache is replaced by a bounded recurrent state, which is what makes long context practical to serve rather than merely supported on paper.

The context window is 262.1K tokens, with up to 131.1K tokens per response. On benchmarks, Qwen3.8 2.4T A95B reaches 86.6 on Terminal Bench 2.1 and leads at 93.0 on PaperBench, while DeepSWE 1.1 at 56.6 and FrontierSWE at 73.5 sit lower relative to the field.

Running the full weights takes serious hardware, at roughly 4.89 TB. Quantized checkpoints bring that down substantially, with 4-bit builds available for both NVIDIA and AMD nodes. Through AI Gateway you skip that entirely and call the model over an API.

You can integrate Qwen3.8 2.4T A95B through AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python.

## What to consider

Qwen3.8 2.4T A95B is the open-weight release, and it is not identical to the hosted Qwen3.8 Max endpoint. The open version omits features the cloud model carries, including image input and a non-thinking mode, so a workload that depends on either belongs on Qwen3.8 Max instead.

Open weights here do not mean unrestricted. The license requires model providers earning more than 50 million US dollars in a twelve-month period to obtain a commercial license from Alibaba Cloud. Read the license terms before you build a hosted product on it.

Benchmark results are uneven by task. Qwen3.8 2.4T A95B leads on PaperBench and posts a strong Terminal Bench 2.1 result, but DeepSWE 1.1 and FrontierSWE land lower. If software engineering is the whole workload, compare it against the alternatives below on your own repository.

## When to use

### Best For

- **Open-weight frontier scale** where you need flagship capability with published weights
- **Long-context work** that a mostly-linear attention stack keeps affordable to serve
- **Research and document tasks** where it leads on PaperBench
- **Terminal and agent workloads** matching its strongest benchmark territory
- **Self-host optionality** where quantized checkpoints let you move off a hosted API later

### Consider Alternatives When

- **Image input** is required, since the open release omits vision that Qwen3.8 Max carries
- **A non-thinking mode** is needed, which the open release also omits
- **Pure software engineering** dominates, where its DeepSWE and FrontierSWE results sit lower
- **Unrestricted licensing** is required, since providers above a revenue threshold need a commercial license

## Best for

- **Open-Weight Frontier Scale:** Flagship capability with published weights
- **Long-Context Work:** A mostly-linear attention stack that stays affordable to serve
- **Research And Document Tasks:** Its leading PaperBench territory
- **Terminal And Agent Workloads:** Where its Terminal Bench result is strongest
- **Self-Host Optionality:** Quantized checkpoints for moving off a hosted API later

## Consider alternatives

- **Image Input Needs:** The open release omits vision that Qwen3.8 Max carries
- **Non-Thinking Mode:** Also omitted from the open release
- **Pure Software Engineering:** DeepSWE 1.1 and FrontierSWE results sit lower
- **Unrestricted Licensing:** Providers above a revenue threshold need a commercial license

## Frequently asked questions

### How is Qwen3.8 2.4T A95B different from Qwen3.8 Max?

Qwen3.8 2.4T A95B is the open-weight release and Qwen3.8 Max is the hosted endpoint. The open version omits some cloud features, including image input and a non-thinking mode.

### How many parameters does Qwen3.8 2.4T A95B use per request?

95 billion active out of 2.4 trillion total, routed across 512 experts. Only the active parameters take part in any single forward pass.

### What is the context window for Qwen3.8 2.4T A95B?

The context window is 262.1K tokens, with up to 131.1K tokens per response. Linear attention on most layers replaces a growing KV cache with a bounded recurrent state, which is what keeps long context servable.

### Is Qwen3.8 2.4T A95B free to use commercially?

The weights are open but the license is not unrestricted. Model providers earning more than 50 million US dollars in a twelve-month period must obtain a commercial license from Alibaba Cloud. Read the license before building a hosted product on it.

### How does Qwen3.8 2.4T A95B perform on coding benchmarks?

Unevenly. It reaches 86.6 on Terminal Bench 2.1, but DeepSWE 1.1 at 56.6 and FrontierSWE at 73.5 sit lower relative to the field. Benchmark it on your own repository before committing a coding workload.

### Do I need my own hardware to run Qwen3.8 2.4T A95B?

Not through AI Gateway. Self-hosting the full weights takes roughly 4.89 TB, though 4-bit quantized checkpoints reduce that substantially for NVIDIA and AMD nodes.

### Does Qwen3.8 2.4T A95B support Zero Data Retention?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/qwen3.8-2.4t-a95b)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)