# Qwen3 Next 80B A3B Instruct

Qwen3 Next 80B A3B Instruct is a Hybrid Transformer-Mamba mixture-of-experts model, 80 billion total parameters with just 3 billion activated per token, designed to deliver high-throughput instruction following over ultra-long context without the computational cost of fully-dense inference.

- **Model ID:** `alibaba/qwen3-next-80b-a3b-instruct`
- **Type:** chat
- **Providers:** vertex, deepinfra
- **Context window:** 262,144
- **Maximum output tokens:** 262,114
- **Pricing:** $0.15/1M input tokens, $1.2/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/qwen3-next-80b-a3b-instruct

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'alibaba/qwen3-next-80b-a3b-instruct',
  prompt: 'Why is the sky blue?'
})
```

## About

Qwen3 Next 80B A3B Instruct introduces a Hybrid Transformer-Mamba architecture that alternates between Gated DeltaNet (a linear attention mechanism) and standard Gated Attention within a 48-layer, 512-expert MoE stack. The layout follows a 12-block repeating pattern: three Gated DeltaNet + MoE layers followed by one Gated Attention + MoE layer. This design is purpose-built for ultra-long-context efficiency: linear attention handles the vast majority of layers at sub-quadratic cost, while sparse Gated Attention layers maintain the precision needed for complex cross-token reasoning.

With only 10 of 512 experts activated per token (plus one shared expert), Qwen3 Next 80B A3B Instruct achieves a sparse activation profile. Combined with Multi-Token Prediction during inference, this delivers notably higher throughput over comparable 32B dense models on sequences of 32K tokens or longer, a meaningful operational advantage for workloads that process long documents or transcripts at scale. See live metrics on this page for current throughput. The Instruct variant is tuned for direct instruction following and doesn't generate thinking traces (that variant is Qwen3-Next-80B-A3B-Thinking).

On the 1M RULER benchmark for extreme-length context, Qwen3 Next 80B A3B Instruct scores 80.3% accuracy, and its context of 262.1K tokens is extensible to approximately one million tokens via YaRN rope scaling. On knowledge benchmarks, it scores 80.6 on MMLU-Pro and 82.7 on Arena-Hard v2, tracking competitively with models that require far more compute per token.

## What to consider

Providers vary in their support for ultra-long context windows; confirm that your selected provider can handle requests approaching 262.1K tokens before deploying at scale.

## When to use

### Best For

- High-throughput document processing pipelines where long context and inference speed must coexist
- Legal, financial, or research workloads that regularly process 100K+ token documents without needing reasoning traces
- Conversational assistants and summarization services that require reliable instruction following over very long sessions
- Multi-document synthesis tasks where entire reports must remain in-context simultaneously
- Cost-sensitive deployments that can't afford dense model pricing at high query volume

### Consider Alternatives When

- If you need explicit chain-of-thought reasoning for hard math or coding problems, use Qwen3-Next-80B-A3B-Thinking instead
- Your task primarily involves short-context exchanges where the architectural advantages for long sequences don't apply
- You require vision or multimodal inputs (this model processes text only)
- Maximum benchmark accuracy on a narrow task takes precedence over throughput efficiency

## Best for

- **High-throughput document pipelines:** Workloads where long context and inference speed must coexist
- **Long-document professional workloads:** Legal, financial, or research tasks that regularly process 100K+ token documents without needing reasoning traces
- **Long-session conversational assistants:** Summarization and chat services that require reliable instruction following across very long sessions
- **Multi-document synthesis:** Tasks where entire reports must remain in-context simultaneously
- **Cost-sensitive deployments:** Production traffic that can't afford dense model pricing at high query volume

## Consider alternatives

- **Explicit chain-of-thought needed:** Hard math or coding problems are better served by Qwen3-Next-80B-A3B-Thinking
- **Short-context exchanges:** The architectural advantages for long sequences don't apply to short prompts
- **Multimodal input required:** This model processes text only; use a vision-language model for images or video
- **Narrow-task accuracy ceiling:** Maximum benchmark accuracy on a specific task can outweigh throughput efficiency gains

## Frequently asked questions

### What does "80B-A3B" mean in this model's name?

"80B" refers to 80 billion total parameters in the MoE pool; "A3B" indicates that approximately 3 billion parameters are activated per token. Only 10 of 512 experts fire for each token, giving the model a sparse activation profile.

### How does Hybrid Transformer-Mamba architecture affect performance on long contexts?

The architecture alternates Gated DeltaNet (linear attention) with sparse Gated Attention. Linear attention scales sub-quadratically with sequence length, enabling the model to process sequences of 262.1K tokens with significantly lower compute than a fully quadratic attention model.

### What is the throughput advantage over a dense model?

On sequences of 32K tokens or longer, Qwen3 Next 80B A3B Instruct achieves notably higher throughput than a comparable Qwen3-32B dense model thanks to its sparse MoE activation and Multi-Token Prediction. See live metrics on this page for current throughput.

### Does this model support a thinking or reasoning mode?

No. The Instruct variant is optimized for direct instruction following without thinking traces. The Qwen3-Next-80B-A3B-Thinking variant provides reasoning mode.

### What is the maximum context length?

The native context is 262.1K tokens. Using YaRN rope scaling, this can be extended to approximately one million tokens. The model achieves 80.3% accuracy on the 1M RULER extreme-context benchmark.

### What benchmarks has this model been evaluated on?

Key scores include MMLU-Pro (80.6), MMLU-Redux (90.9), GPQA (72.9), AIME25 (69.5), LiveCodeBench (56.6), Arena-Hard v2 (82.7), and BFCL-v3 (70.3) for function calling.

### Is Multi-Token Prediction supported?

Yes. Multi-Token Prediction further accelerates inference beyond the baseline throughput gains from the sparse MoE architecture.

## Links

- [Model page](https://vercel.com/ai-gateway/models/qwen3-next-80b-a3b-instruct)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/maas/qwen/qwen3-next-instruct)
- [Provider pricing](https://cloud.google.com/vertex-ai/generative-ai/pricing#qwen-models)