Qwen3.8-Flash-Next README

The project's README from github.com/QwenLM/Qwen3.8-Flash-Next. As one markdown file: README.md.

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:

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

Models

The official model weights are released on:

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.

MLX (Apple Silicon)

If you are running on Apple Silicon, mlx-vlm supports Qwen3.8-Flash-Next (vision + text). Original checkpoints are compatible and can be converted. You can also search for models ending with MLX on the Hugging Face Hub for ready-to-use quantized versions.

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 OpenAI-compatible API service.

sglang serve --model-path Qwen/Qwen3.8-Flash-Next --port 8000 --tp-size 4 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder

An OpenAI-compatible API will be available at http://localhost:8000/v1.

Also see SGLang Cookbook on serving Qwen3.8-Flash-Next.

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM can be used to launch a server with an OpenAI-compatible API service.

vllm serve Qwen/Qwen3.8-Flash-Next --port 8000 --tensor-parallel-size 4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

An OpenAI-compatible API will be available at http://localhost:8000/v1.

Also see vLLM Recipes on serving Qwen3.8-Flash-Next.

TokenSpeed

TokenSpeed is a speed-of-light LLM inference engine. TokenSpeed can be used to launch a server with an OpenAI-compatible API service.

tokenspeed serve Qwen/Qwen3.8-Flash-Next --port 8000 --tensor-parallel-size 4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

An OpenAI-compatible API will be available at http://localhost:8000/v1.

Also see TokenSpeed Recipes on serving Qwen3.8-Flash-Next.

Finetuning

We advise you to use training frameworks, including Unsloth, Swift, Llama-Factory, to finetune your models with SFT, DPO, GRPO, etc.

License Agreement

Please find the license file released with the model weights on Hugging Face Hub or ModelScope.

Citation

If you find our work helpful, feel free to give us a cite.

@techreport{qwen2026design,
    title       = {On the Design of {Qwen3.8-Next} Architecture: Evaluation, Efficiency, and Training Stability},
    author      = {{Qwen Team}},
    institution = {Alibaba Group},
    month       = {August},
    year        = {2026}
}

@misc{qwen3.8flashnext,
    title  = {{Qwen3.8-Flash-Next}: A New Architecture, Towards Ultimate Cost-Efficiency},
    author = {{Qwen Team}},
    month  = {August},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.8-flash-next}
}

Contact Us

If you are interested in leaving a message to either our research team or product team, join our Discord or WeChat groups!

Qwen3.8-Flash-Next (QwenLM/Qwen3.8-Flash-Next)

Searching Database…

433 ·GitHub
Qwen3.8-Flash-Next
What is this?Qwen3.8-Flash-Next is an open-weight multimodal Mixture-of-Experts foundation model from Alibaba's Qwen Team, also serving as an early preview of the architecture Alibaba plans to use for the upcoming Qwen4 model family.
Is this a chatbot app or just raw model weights I have to serve myself?It is just released model weights, not a packaged app. You download them from Hugging Face Hub or ModelScope and run them with an inference framework such as transformers, vLLM, SGLang or llama.cpp.
How do I get it running locally?Download the weights with huggingface download or git clone (or modelscope download), then launch a server with the transformers serve CLI, vLLM, or SGLang to get an OpenAI-compatible API endpoint.
What hardware or framework do I need to run it?You need a serving framework such as transformers, vLLM, SGLang, llama.cpp, MLX (Apple Silicon) or Unsloth, plus enough GPU memory for a 125B-parameter MoE model with 6B parameters active per token.
Can I run the full model on a single consumer GPU?Not in full precision - it's a 125B-parameter MoE model meant to be served through frameworks like vLLM or SGLang on serious hardware, though GGUF quantized builds for llama.cpp and MLX quantized builds for Apple Silicon are available for lighter setups.
What makes the architecture different from earlier Qwen models?It upgrades attention with a Gated DeltaNet + Qwen Sparse Attention hybrid, widens the residual stream into 4 gated branches (Gated Residual), adds an N-gram Embedding table to cheaply scale capacity, and trains with the Muon optimizer for better stability.
Can I try it without hosting it myself?Yes - it powers the new 'Standard' mode on QwenWork, is available through the OpenAI/Anthropic-compatible Qwen API on QwenCloud, and can be used via the Qwen Code terminal agent.
How does it compare to Qwen3.7-Plus?Training costs only about 1/9 as much as Qwen3.7-Plus while delivering superior coding and office-task capabilities.
Can I fine-tune this model myself?Yes, the README recommends training frameworks like Unsloth, Swift, or Llama-Factory to fine-tune it with SFT, DPO, or GRPO.
What are similar projects?Qwen3.8, vllm, Qwen3, unsloth, StartLux-Decision, Hy4-preview, airllm, jeeves
Which topics is it in?dpo, sft, grpo, qwen3, alibaba, open-weights, model, mlx, foundation-models, inference-engine, mixture-of-experts, attention-mechanism
Creator
QwenLM
Size
2.2 MB
Stars
433
Updated
CommitAug 27
Age
1mo
Commits · 90 Days