QwenLM/Qwen3.8-Flash-Next

QwenLM/Qwen3.8-Flash-Next

Description: Qwen3.8-Flash-Next is the foundation model developed by Qwen Team, Alibaba Group.

Stars: 159

Forks: 5

Open issues: 0

Created: 2026-08-24T02:50:39Z

Pushed: 2026-08-26T13:05:07Z

Default branch: main

Fork: no

Archived: no

README:

Qwen3.8-Flash-Next

💜 Qwen Studio | 🤗 Hugging Face | 🤖 ModelScope | 💬 WeChat (微信) | 🫨 Discord

Welcome to the GitHub repository of the Qwen3.8-Flash-Next.

Introduction

Qwen3.8-Flash-Next

In this release we are opening the weights of Qwen3.8-Flash-Next, a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4. It plays the same role that Qwen3-Next played for Qwen3.5: the hybrid Gated DeltaNet + Gated Attention design introduced at that time has since been used across the Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8 series. We are again releasing the architectural changes early, so that the community can examine them before the full Qwen4 model family is built on top of them.

!Qwen3.8-Flash-Next Architecture

Qwen3.8-Flash-Next upgrades the model systematically along four aspects — attention, residual, embedding and optimization — improving model capability while further optimizing computational efficiency, model capacity and training stability:

• Attention A GDN + QSA hybrid architecture. Gated DeltaNet (GDN) compresses the history efficiently; Qwen Sparse Attention (QSA) uses a compressed lightweight indexer to select the important context at micro-block granularity, substantially reducing the cost of attention on long sequences.
• Residual Gated Residual (GR) widens the residual stream into 4 branches and controls reads and writes with a dynamic gate, strengthening cross-layer information flow and training stability.
• Embedding N-gram Embedding looks up a table using the local context to scale model capacity with very little extra computation; the embedding table can be offloaded to host memory and overlapped with model computation through asynchronous prefetching.
• Optimization The Muon optimizer is used, refined around orthogonalization accuracy, the division of labour between Muon and AdamW, and the splitting of fused parameters, with the scaling law refitted for the new architecture.

Qwen3.8-Flash-Next features a 125B-parameter main model, supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. Compared with Qwen3.7-Plus, Qwen3.8-Flash-Next substantially reduces both training and inference cost — training takes only about 1/9 as much, yet it delivers superior capabilities in coding and office tasks.

News

• 2026-08-26: We release Qwen3.8-Flash-Next. Read more on our blog.

Models

The official model weights are released on:

• 🤗Hugging Face Hub: Most LLM frameworks and applications support downloading model files from Hugging Face Hub automatically by specifying the model ID, e.g., Qwen/Qwen3.8-Flash-Next.

You can also download model files manually using huggingface download or git clone. Please follow the instructions on the model page.

• 🤖ModelScope: For users unable to access Hugging Face Hub, we strongly recommend using ModelScope.

For supported frameworks, you can download from ModelScope by setting environment variables, such as SGLANG_USE_MODELSCOPE=true or VLLM_USE_MODELSCOPE=true. You can also download model files manually using modelscope download or git clone. Please follow the instructions on the model page.

Benchmarks

Evaluation results are reported in the Qwen3.8-Flash-Next blog.

Quickstart

Official

You can try Qwen3.8-Flash-Next on our official sites and enjoy the native experience with extra features.

QwenWork

Qwen3.8-Flash-Next now powers the newly-launched "Standard" mode on QwenWork. QwenWork is a one-stop AI working platform launched by Alibaba. Follow its documentation to get started!

Qwen API

QwenCloud provides first-class support for Qwen3.8-Flash-Next, which is compatible with various API specifications, including OpenAI and Anthropic, making it simple for you to try Qwen3.8-Flash-Next in your own applications.

Qwen Code

Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster. Follow its documentation to get started!

Local Use

Hugging Face Transformers

`transformers` acts as the model-definition framework in the current open-weight LLM landscape. It also includes functionalities for LLM inference and training. The addition of serving capabilities in transformers makes it much easier to integrate new models in your development.

To launch a server, simply use the transformers serve command:

transformers serve Qwen/Qwen3.8-Flash-Next --port 8000 --continuous-batching

An OpenAI-compatible API will be available at http://localhost:8000/v1. See the Serve CLI guide for more information.

llama.cpp

`llama.cpp` enables LLM inference with minimal setup and state-of-the-art performance on a wide range of hardware. llama.cpp supports the Qwen3.8-Flash-Next (text & vision). Look for models ending with GGUF on Hugging Face Hub.

Unsloth

Unsloth contains a local UI to run and train LLMs and diffusion models, including Qwen3.8-Flash-Next and more. See the Qwen3.8-Flash-Next guide for running Qwen3.8-Flash-Next quants with Unsloth.

Deployment

Qwen3.8-Flash-Next is supported by multiple inference frameworks. Here we demonstrate the usage of SGLang, vLLM, and TokenSpeed.

SGLang

SGLang is a fast serving framework for large language models and vision language models. SGLang can be used to launch a server with an...