NewChat Models

Alibaba Qwen

Qwen3 Next 80B

Fastest

80B MoE with 3B active parameters and 262K native context. This guide distinguishes the confirmed Instruct and Thinking variants.

80B MoEFast

At a glance

Know the model before you prompt.

Alibaba Qwen API specifications
Context window
262,144 tokens
Maximum output
Not verified
Inputs → output
Text → Text
Knowledge cutoff
Not verified

API model ID: qwen3-next-80b-a3b-instruct / qwen3-next-80b-a3b-thinking

Capabilities & boundaries

What it supports. Where the limits are.

Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.

Supported API features and tools

  • 80B total parameters with 3B activated per token
  • Hybrid attention architecture combining gated attention and Gated DeltaNet
  • 262,144-token native context on both official model cards
  • Separate Instruct and Thinking checkpoints, both Apache 2.0 releases

Before you choose

  • Instruct is non-thinking and Thinking is thinking-only. They are distinct model IDs, not two modes of a single hybrid checkpoint.
  • Context extension up to 1,010,000 requires an appropriate YaRN serving configuration and is not the native limit or an app guarantee.
  • No universal maximum output or hosted token price is established by these cards.
  • The reviewed sources do not establish a knowledge cutoff. Supply current evidence for time-sensitive questions; a large context window is not live web access.

Reasoning behavior

Choose the exact variant: Instruct does not emit thinking blocks; Thinking always reasons and does not support an enable_thinking false shortcut. Use the corresponding chat template and host configuration. Do not silently substitute one for the other.

  • The user-confirmed app IDs map to the official Qwen/Qwen3-Next-80B-A3B-Instruct and Qwen/Qwen3-Next-80B-A3B-Thinking checkpoints.
  • The cards recommend different sampling: Instruct temperature 0.7 / top_p 0.8; Thinking 0.6 / 0.95. These are reference recommendations, not app control guarantees.
  • The Thinking card recommends more than 131,072 context where possible to leave room for reasoning. Memory capacity and hosting settings still determine usable limits.

Put it to work

Start with a more useful prompt.

Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.

Workflow 01

Compare the two confirmed variants

Use identical tasks and a shared rubric.

Design a comparison of the Instruct and Thinking variants on these representative tasks. Define correctness, instruction-following and evidence criteria, and provide a blank table for measured latency and token use. Identify tasks where extra reasoning might help, but do not invent scores or assume one variant always wins.

Workflow 02

Summarize a large technical record

Use a structured evidence index.

Summarize this technical record into confirmed decisions, unresolved issues and next actions. Cite section identifiers for each item and preserve conflicting evidence. Keep the summary concise, mark missing owners as not stated and do not infer that proposals were approved.

Workflow 03

Stress-test a long-context workflow

Check retrieval quality as well as capacity.

Using the supplied document set, design a long-context evaluation with known-answer questions spread across sections. Include distractors, contradictions and missing-answer cases. Define a scoring method and a small-context baseline, leaving performance results blank until the actual endpoint is tested.

Developer reference

Alibaba Qwen API pricing

These are provider API reference prices, not EZ Ai Assist subscription prices.

View EZ Ai Assist plans

No single hosted API price is established by the official open-weight model card. Verify the selected host and model ID; downloadable weights do not make inference free.

  • Self-hosting incurs infrastructure and operating costs. Hosted input, output and cache rates depend on the endpoint; no app charge is inferred from this model card.

Common questions

A few things worth knowing.

Which variant does the app use?

The site owner confirmed both qwen3-next-80b-a3b-instruct and qwen3-next-80b-a3b-thinking. Check the selected app model or route; this single catalog page explains both.

Can a prompt switch Instruct into Thinking?

No. The official cards describe distinct checkpoints. Instruct is non-thinking; Thinking does not support disabling its reasoning with a hybrid-mode flag.

Is 1M the native context?

No. Native context is 262,144. The cards describe optional extension to 1,010,000 with suitable YaRN configuration.

Why is the maximum output unverified?

The reviewed cards do not establish one universal hosted completion ceiling. Generation examples and total context are not substitutes for that limit.

Does this page guarantee access in EZ Ai Assist?

No. A catalog entry and provider capability reference do not establish current app access, enabled tools or available settings. Check the model selector and controls in your workspace.

Are these EZ Ai Assist subscription prices?

No. Provider API rates are developer references. Your subscription and any limits shown in the app are separate; use the website pricing page for plans.

How should I verify an answer?

Provide the relevant evidence and acceptance criteria, ask for source references or reproducible checks, and review the result. Distinguish a proposed tool action from a tool action actually executed.

Check the source

Official documentation

Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.

Same provider