Qwen3-Next: Models, 256K Context & Local Setup

Qwen3-Next is an open-weight, text-only mixture-of-experts model family. Its public release uses separate Instruct and Thinking checkpoints. It is not one downloadable checkpoint that switches dynamically between the two behaviors.

Last verified: July 21, 2026.

Checkpoint rule: Select Qwen/Qwen3-Next-80B-A3B-Instruct for non-thinking instruction following or Qwen/Qwen3-Next-80B-A3B-Thinking for thinking-mode output. Do not add a fictional runtime toggle.

Official Qwen3-Next checkpoints

Model ID	Behavior	Format
`Qwen/Qwen3-Next-80B-A3B-Instruct`	Instruct/non-thinking only	BF16 weights
`Qwen/Qwen3-Next-80B-A3B-Thinking`	Thinking only	BF16 weights
`Qwen/Qwen3-Next-80B-A3B-Instruct-FP8`	Instruct/non-thinking only	FP8 weights
`Qwen/Qwen3-Next-80B-A3B-Thinking-FP8`	Thinking only	FP8 weights

The Instruct checkpoint does not generate <think></think> blocks. The Thinking checkpoint’s chat template enforces thinking. Its decoded output can contain a closing </think> marker without a visible opening marker because the opening marker may be inserted by the template.

Architecture and parameter count

Model type	Text-only causal language model
Total parameters	80B
Activated parameters per token	3B
Architecture	Sparse MoE with Gated DeltaNet and Gated Attention
Native context	262,144 tokens
Extended context	Up to approximately 1M with YaRN configuration
License	Apache License 2.0

The “3B activated” figure does not mean the downloaded checkpoint contains only 3B parameters or has the memory requirements of a dense 3B model. The published model contains 80B total parameters and activates a subset for each token.

Native 256K context versus extended 1M context

Qwen3-Next natively supports 262,144 tokens. Qwen also documents validation up to approximately one million tokens by applying YaRN scaling, with an extended serving value of 1,010,000 tokens.

The one-million-token setting is an extrapolated configuration rather than the native context. Qwen warns that static YaRN can affect shorter-context performance, so it should be enabled only when the application needs context beyond the native limit.

Which checkpoint should you choose?

Choose Instruct for concise assistant responses, summarization, extraction, ordinary coding help, or predictable non-thinking output.
Choose Thinking for difficult mathematics, multi-step reasoning, complex debugging, or tasks where additional reasoning tokens are acceptable.
Choose an FP8 repository only when your serving framework and accelerator configuration support that format.

Transformers example

Use a Transformers version that includes Qwen3-Next support.

pip install "transformers>=4.57.1" accelerate

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen3-Next-80B-A3B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": (
            "Explain the difference between native and "
            "extrapolated context in two paragraphs."
        ),
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(
    [prompt],
    return_tensors="pt",
).to(model.device)

generated = model.generate(
    **inputs,
    max_new_tokens=512,
)

new_tokens = generated[0][inputs.input_ids.shape[-1]:]

answer = tokenizer.decode(
    new_tokens,
    skip_special_tokens=True,
)

print(answer)

To use the reasoning checkpoint, change the model ID to Qwen/Qwen3-Next-80B-A3B-Thinking. No enable_thinking parameter is required or documented for switching the Instruct checkpoint.

Serving Qwen3-Next with vLLM

Qwen documents vLLM 0.10.2 or later for Qwen3-Next. The following command reproduces the official four-way tensor-parallel example; it is an example configuration, not a universal minimum-hardware requirement.

pip install "vllm>=0.10.2"

vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct \
  --port 8000 \
  --tensor-parallel-size 4 \
  --max-model-len 262144

Reduce --max-model-len if the serving process cannot reserve enough memory. The allocated context length can have a major effect on memory use even when individual requests are short.

Tool-using applications

Qwen documents agent workflows through Qwen-Agent and compatible serving frameworks. In these systems, the model can propose a tool call, but the application defines the tools, validates arguments, controls permissions, performs execution, and returns results.

Limitations

Qwen3-Next accepts text and produces text; it is not a vision, audio, or video model.
One million tokens requires YaRN and is not the native context setting.
The 80B checkpoint should not be described as having the deployment cost of a dense 3B model.
Memory and throughput depend on precision, context allocation, framework, and hardware.
Tool execution must remain under application control.

Frequently asked questions

Can Qwen3-Next switch between Thinking and Instruct?

Not within one published checkpoint. Qwen provides separate Instruct and Thinking model repositories.

Does Qwen3-Next have a native 1M context?

No. Its native context is 262,144 tokens. Approximately one million tokens requires YaRN configuration.

Does Qwen3-Next understand images or video?

No. The published Qwen3-Next checkpoints are text-only causal language models.

Is Qwen3-Next open weight?

Yes. The listed repositories use the Apache License 2.0.

Official references

This is an independent technical guide and is not affiliated with or endorsed by Alibaba Cloud or the Qwen team.

Qwen3-Next: Instruct vs Thinking, Context, and Setup