Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is Qwen's open multimodal Mixture-of-Experts model and early preview of the architecture planned for Qwen4, with 6B active parameters, long context, and official local or hosted serving paths.

Language ModelOpen WeightsMultimodalMixture of ExpertsLong ContextReasoning ModelQwen

Self-managed Transformers, llama.cpp, SGLang, vLLM, or TokenSpeed inference; QwenWork and QwenCloud are separate managed surfaces

Open weights have no provider token price; hosted pricing and model identifiers are provider-specific, and commercial use is subject to Qwen Community License 1.0

Commercial

🚀Function Overview

An experimental open-weight vision-language model for evaluating Qwen's next attention, residual, embedding, and optimization design on coding, office, agent, long-context, and multimodal work.

Key Features

Official checkpoint ID Qwen/Qwen3.8-Flash-Next
Official FP8 checkpoint ID Qwen/Qwen3.8-Flash-Next-FP8
125B-parameter main model with 6B activated per token, plus 51B n-gram embedding parameters and 4B MTP parameters
Hybrid Gated DeltaNet and Qwen Sparse Attention architecture
Text, image, and video input with text output
262,144-token native context extensible to 1,000,000 tokens with YaRN
Thinking enabled by default but configurable for direct responses
Official Transformers, llama.cpp, SGLang, vLLM, TokenSpeed, QwenWork, and QwenCloud paths
Qwen Community License 1.0 with additional commercial conditions
Experimental preview of architectural ideas planned for Qwen4, not a released Qwen4 model

Use Cases

•Evaluate long-context coding and repository tasks with sparse attention
•Test office, tool-use, and agent workflows with optional reasoning
•Analyze images, video, documents, and screens with a self-hosted multimodal model
•Research Qwen Sparse Attention, gated residual streams, and offloadable n-gram embeddings
•Compare open self-hosting with the separate hosted Qwen3.8-Flash production surface

⚙️Input Parameters

model

string

Use Qwen/Qwen3.8-Flash-Next, Qwen/Qwen3.8-Flash-Next-FP8, or the exact hosted model identifier documented by the selected provider.

messages

array

Conversation messages containing text and supported image or video content.

enable_thinking

boolean

Enable or disable thinking when the selected API or serving framework supports this control. Thinking is enabled by default.

max_tokens

integer

Set an output allowance appropriate to the task and serving path. Qwen recommends up to 262,144 reasoning tokens and 131,072 final-response tokens within a 1M context configuration.

media

array

Optional image or video inputs. Video sampling controls vary by serving framework.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "Qwen/Qwen3.8-Flash-Next",
  "messages": [
    {
      "role": "user",
      "content": "Review this architecture diagram and migration plan, then identify the highest-risk assumption and the smallest validation sequence."
    }
  ],
  "enable_thinking": true,
  "max_tokens": 8192
}

Output Results

A multimodal chat completion from the self-hosted or managed Qwen3.8-Flash-Next endpoint. Verify reasoning schema, latency, memory, tool behavior, and license fit on the selected serving path.

Quick Actions

Technical Specifications

Hardware Type: Self-managed Transformers, llama.cpp, SGLang, vLLM, or TokenSpeed inference; QwenWork and QwenCloud are separate managed surfaces
Commercial Use: Supported
Pricing: Open weights have no provider token price; hosted pricing and model identifiers are provider-specific, and commercial use is subject to Qwen Community License 1.0

Related Keywords

Qwen3.8-Flash-NextQwen 3.8 Flash NextQwen3.8 Flash Next open weightsQwen3.8 Flash Next licenseQwen3.8 Flash Next context lengthQwen3.8 Flash Next local deploymentQwen3.8 Flash Next vLLMQwen3.8 Flash Next SGLangQwen3.8 Flash Next GGUFQwen4 architecture preview

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