Qwen3-Next: Towards Ultimate Training & Inference Efficiency

Introduction#

We believe that Context Length Scaling and Total Parameter Scaling are two major trends in the future of large models. To further improve training and inference efficiency under long-context and large-parameter settings, we design a brand-new model architecture called Qwen3-Next.Compared to the MoE structure of Qwen3, Qwen3-Next introduces several key improvements: a hybrid attention mechanism, a highly sparse Mixture-of-Experts (MoE) structure, training-stability-friendly optimizations, and a multi-token prediction mechanism for faster inference.

Based on this new architecture, we train the Qwen3-Next-80B-A3B-Base model — an 80-billion-parameter model that activates only 3 billion parameters during inference. This base model achieves performance comparable to (or even slightly better than) the dense Qwen3-32B model, while using less than 10% of its training cost (GPU hours). In inference, especially with context lengths over 32K tokens, it delivers more than 10x higher throughput — achieving extreme efficiency in both training and inference.

We develop and release two post-trained versions based on Qwen3-Next-80B-A3B-Base: Qwen3-Next-80B-A3B-Instruct and Qwen3-Next-80B-A3B-Thinking. We solve the long-standing stability and efficiency issues in reinforcement learning (RL) training caused by the hybrid attention + high-sparsity MoE architecture. This led to improvements in both RL training speed and final performance.

The Qwen3-Next-80B-A3B-Instruct performs comparably to our flagship model Qwen3-235B-A22B-Instruct-2507, and shows clear advantages in tasks requiring ultra-long context (up to 256K tokens). The Qwen3-Next-80B-A3B-Thinking excels at complex reasoning tasks — outperforming higher-cost models like Qwen3-30B-A3B-Thinking-2507 and Qwen3-32B-Thinking, outpeforming the closed-source Gemini-2.5-Flash-Thinking on multiple benchmarks, and approaching the performance of our top-tier model Qwen3-235B-A22B-Thinking-2507.

We have already released Qwen3-Next on Hugging Face and ModelScope. Additionally, everyone can use the Qwen3-Next service via Alibaba Cloud Model Studio and NVIDIA API Catalog.

Key Features#

Hybrid Architecture: Gated DeltaNet + Gated Attention

Linear attention breaks the quadratic complexity of standard attention and is more efficient for long contexts. However, we find that using only linear attention or only standard attention has limitations: linear attention is fast but weak at recall; standard attention is expensive and slow during inference.

Through systematic experiments, we find that Gated DeltaNet [1] offers stronger in-context learning ability than commonly used methods like Sliding Window Attention or Mamba2. When we mix Gated DeltaNet with standard attention at a 3:1 ratio (75% layers use Gated DeltaNet, 25% keep standard attention), the model consistently outperforms any monolithic architecture — achieving both better performance and higher efficiency. In the standard attention layers, we add several enhancements:

Adopt the output gating mechanism from our prior work [2] to reduce low-rank issues in attention.
Increase the dimension per attention head from 128 to 256.
Apply rotary position encoding only to the first 25% of position dimensions — improving extrapolation to longer sequences.

Ultra-Sparse MoE: Activating Only 3.7% of Parameters

Qwen3-Next uses a highly sparse MoE design: 80B total parameters, but only ~3B activated per inference step. Experiments show that, with global load balancing [4], increasing total expert parameters while keeping activated experts fixed steadily reduces training loss.

Compared to Qwen3’s MoE (128 total experts, 8 routed), Qwen3-Next expands to 512 total experts, combining 10 routed experts + 1 shared expert — maximizing resource usage without hurting performance.

Training-Stability-Friendly Designs

We found that the attention output gating mechanism helps eliminate issues like Attention Sink [5] and Massive Activation [6] , ensuring numerical stability across the model.

In Qwen3, we use QK-Norm, but notice some layer norm weights become abnormally large. To fix this and further improve stability, Qwen3-Next adopts Zero-Centered RMSNorm [7], and applies weight decay to norm weights to prevent unbounded growth.

We also normalize MoE router parameters during initialization [8] , ensuring each expert is unbiasedly selected early in training — reducing noise from random initialization.

These stability-focused designs make small-scale experiments more reliable and help large-scale training run smoothly.

Multi-Token Prediction

Qwen3-Next introduces a native Multi-Token Prediction (MTP) [3][9][10] mechanism, which not only yields an MTP module with a high acceptance rate for Speculative Decoding but also enhances the overall performance.Additionally, Qwen3-Next specifically optimizes the multi-step inference performance of MTP, further improving the acceptance rate of Speculative Decoding in real scenarios through multi-step training that maintains consistency between training and inference.

Pre-training#

Pretraining Efficiency & Inference Speed

Qwen3-Next is trained on a uniformly sampled subset (15T tokens) of Qwen3’s 36T-token pretraining corpus. It uses less than 80% of the GPU hours needed by Qwen3-30A-3B, and only 9.3% of the compute cost of Qwen3-32B — while achieving better performance. This shows outstanding training efficiency and value.

Thanks to its hybrid architecture, Qwen3-Next also shines in inference:

Prefill Stage: At 4K context length, throughput is nearly 7x higher than Qwen3-32B. Beyond 32K, it’s over 10x faster.
Decode Stage: At 4K context, throughput is nearly 4x higher. Even beyond 32K, it still maintains over 10x speed advantage.

Base Model Performance

Qwen3-Next-80B-A3B-Base activates only 1/10th of the non-embedding parameters used by Qwen3-32B-Base — yet outperforms it on most benchmarks, and significantly beats Qwen3-30B-A3B. This demonstrates exceptional efficiency and strong performance.

Post-training#

Instruct Model Performance

Qwen3-Next-80B-A3B-Instruct significantly outperforms Qwen3-30B-A3B-Instruct-2507 and Qwen3-32B-Non-thinking, and achieves results nearly matching our flagship Qwen3-235B-A22B-Instruct-2507.

On RULER, Qwen3-Next-80B-A3B-Instruct outperforms Qwen3-30B-A3B-Instruct-2507 (which has more attention layers) across all lengths — and even beats Qwen3-235B-A22B-Instruct-2507 (which has more layers overall) within 256K context. This proves the strength of the Gated DeltaNet + Gated Attention hybrid design for long-context tasks.

Thinking Model Performance

Qwen3-Next-80B-A3B-Thinking outperforms higher-cost models like Qwen3-30B-A3B-Thinking-2507 and Qwen3-32B-Thinking. It beats the closed-source Gemini-2.5-Flash-Thinking on multiple benchmarks, and comes close to our latest flagship model Qwen3-235B-A22B-Thinking-2507 on key metrics.

Develop with Qwen3#

All examples below are based on the Qwen3-Next-80B-A3B-Instruct version. For the Thinking model, please refer to https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking.

Hugging Face Transformers#

The code for Qwen3-Next has been merged into the main branch of Hugging Face transformers.}}

With earlier versions, you will encounter the following error:}}

The following contains a code snippet illustrating how to use the model generate content based on given inputs. }}

Note:

Multi-Token Prediction (MTP) is not generally available in Hugging Face Transformers.

Note:

The efficiency or throughput improvement depends highly on the implementation.

It is recommended to adopt a dedicated inference framework, e.g., SGLang and vLLM, for inference tasks.

Tip

Depending on the inference settings, you may observe better efficiency with flash-linear-attention and causal-conv1d.

See the above links for detailed instructions and requirements.

For deployment, you can use the latest sglang or vllm to create an OpenAI-compatible API endpoint.

SGLang#

SGLang is a fast serving framework for large language models and vision language models. SGLang could be used to launch a server with OpenAI-compatible API service.

SGLang has supported Qwen3-Next in its main branch, which can be installed from source:}}

The following command can be used to create an API endpoint at http://localhost:30000/v1 with maximum context length 256K tokens using tensor parallel on 4 GPUs.}}

The following command is recommended for MTP with the rest settings the same as above:}}

Note

The environment variable SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 is required at the moment.

Note

The default context length is 256K. Consider reducing the context length to a smaller value, e.g., 32768, if the server fail to start.

vLLM#

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

vLLM has supported Qwen3-Next in its main branch, which can be installed from source:}}

The following command can be used to create an API endpoint at http://localhost:8000/v1 with maximum context length 256K tokens using tensor parallel on 4 GPUs.}}

The following command is recommended for MTP with the rest settings the same as above:}}

Note

The environment variable VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 is required at the moment.

Note

The default context length is 256K. Consider reducing the context length to a smaller value, e.g., 32768, if the server fail to start.

Agentic Use#

Qwen3 excels in tool calling capabilities. We recommend using Qwen-Agent to make the best use of agentic ability of Qwen3. Qwen-Agent encapsulates tool-calling templates and tool-calling parsers internally, greatly reducing coding complexity.

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.}}

Processing Ultra-Long Texts#

Qwen3-Next natively supports context lengths of up to 262,144 tokens. For conversations where the total length (including both input and output) significantly exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively. We have validated the model's performance on context lengths of up to 1 million tokens using the YaRN method.

YaRN is currently supported by several inference frameworks, e.g., transformers, vllm and sglang. In general, there are two approaches to enabling YaRN for supported frameworks:

Modifying the model files: In the config.json file, add the rope_scaling fields:
Passing command line arguments:

For vllm, you can use

For sglang, you can use

NOTE:

All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts.

We advise adding the rope_scaling configuration only when processing long contexts is required.

It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.

Summary#

Qwen3-Next represents a major leap forward in model architecture, introducing innovations in attention mechanisms, including linear attention and attention gate, as well as increased sparsity in its MoE design. Qwen3-Next-80B-A3B delivers performance on par with the larger Qwen3-235B-A22B-2507 across both thinking and non-thinking modes, while offering significantly faster inference, especially in long-context scenarios. With this release, we aim to empower the open-source community to evolve alongside cutting-edge architectural advances. Looking ahead, we will further refine this architecture to develop Qwen3.5, targeting unprecedented levels of intelligence and productivity.

References#

[^1]: Gated Delta Networks: Improving Mamba2 with Delta Rule

[^2]: Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free

[^3]: DeepSeek-V3 Technical Report

[^4]: Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models

[^5]: Efficient Streaming Language Models with Attention Sinks

[^6]: Massive Activations in Large Language Models

[^7]: Gemma: Open Models Based on Gemini Research and Technology

[^8]: Approximating Two-Layer Feedforward Networks for Efficient Transformers

[^9]: Better & faster large language models via multi-token prediction

[^10]: ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training