Qwen3-Next-80B-A3B-Instruct

Over the past few months, we have observed increasingly clear trends toward scaling both total parameters and context lengths in the pursuit of more powerful and agentic artificial intelligence (AI). We are excited to share our latest advancements in addressing these demands, centered on improving scaling efficiency through innovative model architecture. We call this next-generation foundation models Qwen3-Next.

Highlights

Qwen3-Next-80B-A3B is the first installment in the Qwen3-Next series and features the following key enchancements:

• Hybrid Attention: Replaces standard attention with the combination of Gated DeltaNet and Gated Attention, enabling efficient context modeling for ultra-long context length.
• High-Sparsity Mixture-of-Experts (MoE): Achieves an extreme low activation ratio in MoE layers, drastically reducing FLOPs per token while preserving model capacity.
• Stability Optimizations: Includes techniques such as zero-centered and weight-decayed layernorm, and other stabilizing enhancements for robust pre-training and post-training.
• Multi-Token Prediction (MTP): Boosts pretraining model performance and accelerates inference.

We are seeing strong performance in terms of both parameter efficiency and inference speed for Qwen3-Next-80B-A3B:

• Qwen3-Next-80B-A3B-Base outperforms Qwen3-32B-Base on downstream tasks with 10% of the total training cost and with 10 times inference throughput for context over 32K tokens.
• Qwen3-Next-80B-A3B-Instruct performs on par with Qwen3-235B-A22B-Instruct-2507 on certain benchmarks, while demonstrating significant advantages in handling ultra-long-context tasks up to 256K tokens.

For more details, please refer to our blog post Qwen3-Next.

Model Overview

[!Note] Qwen3-Next-80B-A3B-Instruct supports only instruct (non-thinking) mode and does not generate <think></think> blocks in its output.

Qwen3-Next-80B-A3B-Instruct has the following features:

• Type: Causal Language Models
• Training Stage: Pretraining (15T tokens) & Post-training
• Number of Parameters: 80B in total and 3B activated
• Number of Parameters (Non-Embedding): 79B
• Number of Layers: 48
• Hidden Dimension: 2048
• Hybrid Layout: 12 * (3 * (Gated DeltaNet -> MoE) -> (Gated Attention -> MoE))
• Gated Attention:
  • Number of Attention Heads: 16 for Q and 2 for KV
  • Head Dimension: 256
  • Rotary Position Embedding Dimension: 64
• Gated DeltaNet:
  • Number of Linear Attention Heads: 32 for V and 16 for QK
  • Head Dimension: 128
• Mixture of Experts:
  • Number of Experts: 512
  • Number of Activated Experts: 10
  • Number of Shared Experts: 1
  • Expert Intermediate Dimension: 512
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens

Performance

| | Qwen3-30B-A3B-Instruct-2507 | Qwen3-32B Non-Thinking | Qwen3-235B-A22B-Instruct-2507 | Qwen3-Next-80B-A3B-Instruct | |--- | --- | --- | --- | --- | | Knowledge | | | | | | MMLU-Pro | 78.4 | 71.9 | 83.0 | 80.6 | | MMLU-Redux | 89.3 | 85.7 | 93.1 | 90.9 | | GPQA | 70.4 | 54.6 | 77.5 | 72.9 | | SuperGPQA | 53.4 | 43.2 | 62.6 | 58.8 | | Reasoning | | | | | | AIME25 | 61.3 | 20.2 | 70.3 | 69.5 | | HMMT25 | 43.0 | 9.8 | 55.4 | 54.1 | | LiveBench 20241125 | 69.0 | 59.8 | 75.4 | 75.8 | | Coding | | | | | | LiveCodeBench v6 (25.02-25.05) | 43.2 | 29.1 | 51.8 | 56.6 | | MultiPL-E | 83.8 | 76.9 | 87.9 | 87.8 | | Aider-Polyglot | 35.6 | 40.0 | 57.3 | 49.8 | | Alignment | | | | | | IFEval | 84.7 | 83.2 | 88.7 | 87.6 | | Arena-Hard v2* | 69.0 | 34.1 | 79.2 | 82.7 | | Creative Writing v3 | 86.0 | 78.3 | 87.5 | 85.3 | | WritingBench | 85.5 | 75.4 | 85.2 | 87.3 | | Agent | | | | | | BFCL-v3 | 65.1 | 63.0 | 70.9 | 70.3 | | TAU1-Retail | 59.1 | 40.1 | 71.3 | 60.9 | | TAU1-Airline | 40.0 | 17.0 | 44.0 | 44.0 | | TAU2-Retail | 57.0 | 48.8 | 74.6 | 57.3 | | TAU2-Airline | 38.0 | 24.0 | 50.0 | 45.5 | | TAU2-Telecom | 12.3 | 24.6 | 32.5 | 13.2 | | Multilingualism | | | | | | MultiIF | 67.9 | 70.7 | 77.5 | 75.8 | | MMLU-ProX | 72.0 | 69.3 | 79.4 | 76.7 | | INCLUDE | 71.9 | 70.9 | 79.5 | 78.9 | | PolyMATH | 43.1 | 22.5 | 50.2 | 45.9 |

*: For reproducibility, we report the win rates evaluated by GPT-4.1.

Quickstart

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.

Deployment

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:

  JSONCopy

• Passing command line arguments:

  For vllm, you can use

  ShellCopy

  For sglang, you can use

  ShellCopy

[!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.

Long-Context Performance

We test the model on an 1M version of the RULER benchmark.

| Model Name | Acc avg | 4k | 8k | 16k | 32k | 64k | 96k | 128k | 192k | 256k | 384k | 512k | 640k | 768k | 896k | 1000k | |---------------------------------------------|---------|------|------|------|------|------|------|------|------|------|------|------|------|------|------|-------| | Qwen3-30B-A3B-Instruct-2507 | 86.8 | 98.0 | 96.7 | 96.9 | 97.2 | 93.4 | 91.0 | 89.1 | 89.8 | 82.5 | 83.6 | 78.4 | 79.7 | 77.6 | 75.7 | 72.8 | | Qwen3-235B-A22B-Instruct-2507 | 92.5 | 98.5 | 97.6 | 96.9 | 97.3 | 95.8 | 94.9 | 93.9 | 94.5 | 91.0 | 92.2 | 90.9 | 87.8 | 84.8 | 86.5 | 84.5 | | Qwen3-Next-80B-A3B-Instruct | 91.8 | 98.5 | 99.0 | 98.0 | 98.7 | 97.6 | 95.0 | 96.0 | 94.0 | 93.5 | 91.7 | 86.9 | 85.5 | 81.7 | 80.3 | 80.3 |

• Qwen3-Next are evaluated with YaRN enabled. Qwen3-2507 models are evaluated with Dual Chunk Attention enabled.
• Since the evaluation is time-consuming, we use 260 samples for each length (13 sub-tasks, 20 samples for each).

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:

   • We suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=0.
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

2. Adequate Output Length: We recommend using an output length of 16,384 tokens for most queries, which is adequate for instruct models.

3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.

   • Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
   • Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

Citation

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