LLMPrivate

Qwen 3 Next 80b

Alibaba's 80B/3B sparse MoE with hybrid attention and 256K context, open-weight under Apache 2.0.

Maker
Alibaba / Qwen team
Modality
Text
Context
256K tokens
Open weights
Yes — Apache 2.0

Overview

What is Qwen 3 Next 80b

Qwen 3 Next 80B is Alibaba's open-weight sparse Mixture-of-Experts language model with 80 billion total parameters and 3 billion active per token. Released in September 2025, it uses hybrid attention and multi-token prediction to deliver high throughput across contexts up to 256K tokens.

Running it privately on Venice

Run Qwen 3 Next 80B on Venice with zero retention — your prompts are not stored or used for training. Venice hosts the open weights directly with tool use, web search, and structured JSON output available, giving you permissionless sovereignty over your inference without Big-Tech surveillance.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Extreme parameter efficiency: 80B total parameters with only 3B active per token, outperforming dense Qwen3-32B at under one-tenth the training cost.
  • Ultra-long context performance: hybrid attention enables 10× higher throughput than dense models beyond 32K tokens and strong results up to 256K tokens.
  • Fully open weights released under Apache 2.0, enabling self-hosting, fine-tuning, and auditability.
  • On Venice it supports tool use, web search, and structured JSON schema output for agentic and automation workflows.
  • Benchmarked on par with Qwen3-235B-A22B-Instruct on certain tasks despite a fraction of the active parameters.
Limitations
  • MoE architectures can exhibit higher latency for single-token streaming compared to similarly priced dense models.
  • Instruction-following on precise multi-step prompts may trail top-tier closed frontier models according to third-party evaluations.
  • Output token pricing is steeper than some open-weight rivals, making high-volume generation more expensive.
  • Not an uncensored model: safety filters and standard refusals still apply.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
Alibaba / Qwen team
Released
September 2025
Modality
Text generation
Architecture
Sparse MoE with Hybrid Attention (Gated DeltaNet + Gated Attention) and Multi-Token Prediction
Parameters
80B total / 3B active per token
Open weights
Yes — Apache 2.0
Context window
256K tokens
Max output
16.384K tokens
Capabilities
Function calling, Web search
Privacy on Venice
Private — zero retention
Available on Venice since
Apr 2025

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-next-80b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.35 per 1M input tokens and $1.90 per 1M output tokens.

Input / 1M tokens
$0.35
Per 1M tokens
Output / 1M tokens
$1.90
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelContext windowOpen weightsPrice (Venice)Best for
Qwen 3 Next 80B256K tokensYes$0.35 in · $1.90 out / 1MLong-context efficiency
DeepSeek V3.2160K tokensYes$0.33 in · $0.48 out / 1MBudget inference
Kimi K2.6256K tokensYes$0.75 in · $3.50 out / 1MDense-model quality
Claude Sonnet 4.61M tokensNo$3.60 in · $18 out / 1MClosed frontier

Sparse 80B/3B MoE with hybrid attention — extreme throughput for 256K contexts at a mid-tier price.

Use cases

What it is good for

  1. 01Long-document analysis, summarization, and RAG across 256K token contexts.
  2. 02Agentic automation with function calling, web search, and structured JSON output.
  3. 03Code generation and repository-level understanding leveraging the large context window.
  4. 04Cost-efficient large-scale inference where open weights and data sovereignty are required.
  5. 05High-throughput batch processing of long-context tasks thanks to sparse activation.

Prompting

Getting better results

Feed full documents or extended conversation histories into a single prompt to exploit the 256K context window.

Use structured JSON schema output for reliable data extraction and downstream pipeline integration.

For coding, include relevant file context and imports to help the model reason across long dependencies.

FAQ

Frequently asked questions

Qwen 3 Next 80B is Alibaba's open-weight sparse Mixture-of-Experts language model with 80 billion total parameters and 3 billion active per token. Released in September 2025, it uses hybrid attention and multi-token prediction to deliver high throughput across contexts up to 256K tokens.

On Venice, Qwen 3 Next 80B is billed at $0.35 per 1M input tokens and $1.90 per 1M output tokens. You pay only for the tokens you use with no subscription required.

Yes. The weights are released under the Apache 2.0 license on Hugging Face, Kaggle, and ModelScope. You can self-host, fine-tune, or run it privately through Venice's hosted inference.

Yes. On Venice, the model supports function calling, web search, and structured JSON schema output, making it suitable for agentic workflows and data extraction.

Qwen 3 Next 80B offers a larger 256K context window and a sparse 80B/3B MoE architecture optimized for ultra-long contexts. DeepSeek V3.2 is cheaper per token but uses a dense architecture and a 160K context window. Choose Qwen for long-document workloads and DeepSeek for cost-sensitive, shorter-context tasks.

It uses a sparse Mixture-of-Experts design with hybrid attention combining Gated DeltaNet and Gated Attention, plus Multi-Token Prediction. Only 3 billion parameters are activated per token, enabling high efficiency despite the 80 billion parameter total.

Venice runs Qwen 3 Next 80B under a private zero-retention tier. Your prompts are not stored, profiled, or used for training, and no account history is built from your generations.

Run Qwen 3 Next 80b privately

No prompt logging. No data used for training.