Qwen3.8-2.4T-A95B

Qwen3.8-2.4T-A95B is Qwen's open-weight, text-only Mixture-of-Experts model with 2.4 trillion total parameters, 95 billion activated per token, mandatory thinking, and official distributed-serving recipes.

Platform: Replicate

Language ModelOpen WeightsMixture of ExpertsLong ContextReasoning ModelQwen

Distributed multi-node accelerators through SGLang, vLLM, or TokenSpeed

Open weights have no provider token price; infrastructure cost varies, and commercial use is subject to the Qwen3.8-Max License

Commercial

🚀Function Overview

A very large open-weight reasoning model for distributed self-hosting and controlled evaluation. Let's explore what this AI model can do for you!

Key Features

Official checkpoint ID Qwen/Qwen3.8-2.4T-A95B
Official FP8 checkpoint ID Qwen/Qwen3.8-2.4T-A95B-FP8
2.4 trillion total parameters with 95 billion activated per token
512 experts with 10 routed experts plus one shared expert activated
262,144-token native context length extensible to 1,010,000 tokens
Text input and text output only; multimodal input is not supported
Thinking mode is required and cannot be disabled
Low, medium, and xhigh reasoning effort, with xhigh documented as the default
Official SGLang, vLLM, and TokenSpeed serving recipes
Custom Qwen3.8-Max License with commercial attribution and separate-license conditions

Use Cases

•Distributed evaluation of Qwen's largest open model on coding and agent tasks
•Self-hosted text reasoning where infrastructure, data control, or sustained volume justifies the operating cost
•Research on sparse Mixture-of-Experts routing and long-context behavior
•Fine-tuning or derivative-work experiments allowed by the Qwen3.8-Max License
•Comparing the open checkpoint with the hosted Qwen3.8-Max API without assuming feature parity

⚙️Input Parameters

model

string

Use Qwen/Qwen3.8-2.4T-A95B, Qwen/Qwen3.8-2.4T-A95B-FP8, or the exact local model path served by your inference framework.

messages

array

Text-only conversation messages. The open checkpoint does not accept image or video inputs.

reasoning_effort

string

Optional reasoning-depth control: low, medium, or xhigh. The model card documents xhigh as the default.

preserve_thinking

boolean

Optional control for retaining reasoning context from earlier messages when the serving framework supports it.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "Qwen/Qwen3.8-2.4T-A95B",
  "messages": [
    {
      "role": "user",
      "content": "Review this migration plan, identify the riskiest assumption, and propose a bounded validation sequence."
    }
  ],
  "reasoning_effort": "medium"
}

Output Results

A self-hosted chat completion whose response begins with thinking content before the final answer. Verify the exact schema and reasoning parser in the selected serving framework.

Quick Actions

Use NowView Documentation

Technical Specifications

Hardware Type: Distributed multi-node accelerators through SGLang, vLLM, or TokenSpeed
Commercial Use: Supported
Pricing: Open weights have no provider token price; infrastructure cost varies, and commercial use is subject to the Qwen3.8-Max License
Platform: Replicate

Related Keywords

Qwen3.8-2.4T-A95BQwen 3.8 open weightsQwen3.8 FP8Qwen3.8 licenseQwen3.8 local deploymentQwen3.8 vLLMQwen3.8 SGLangQwen3.8 95B activeQwen3.8 context length

Related Models

Bielik 1.5B v3 Instruct

Bielik-1.5B-v3-Instruct is a generative text model featuring 1.6 billion parameters. It is result of collaboration between the open-science/open-souce project SpeakLeash and the High Performance Computing (HPC)

Cordia-A6 Text Generation Model

A model for generating text sequences based on input prompts and adjustable parameters.

Claude Sonnet 4

Claude Sonnet 4 is a significant upgrade to 3.7, delivering superior coding and reasoning while responding more precisely to your instructions