Qwen3-Coder-Next

State-of-the-art coding agent with ultra-efficient 3B active inference.

Try Now

read docs

About model

Qwen3-Coder-Next is an open-weight language model designed specifically for coding agents. With only 3B activated parameters (80B total), it achieves performance comparable to models with 10–20x more active parameters, making it highly cost-effective for production agent deployment. Through an elaborate training recipe, Qwen3-Coder-Next excels at long-horizon reasoning, complex tool usage, and recovery from execution failures, ensuring robust performance in dynamic coding tasks. With 256K context length and advanced tool calling capabilities, it delivers state-of-the-art agentic coding on Together AI's production infrastructure.

SWE-Bench Verified (w/ SWE-Agent)

74.2%

Production-level autonomous coding

Activated Parameters

Performing like 30-60B models

Parameter Efficiency

10-20x

Cost savings for agent workloads

Model key capabilities

Ultra-Efficient Architecture: Only 3B activated parameters (80B total) achieving performance comparable to models with 10-20x more active parameters—highly cost-effective for production agent deployment
Advanced Agentic Capabilities: Long-horizon reasoning, complex tool usage, and recovery from execution failures—ensuring robust performance in dynamic coding workflows
Leading Coding Performance: 74.2% SWE-Bench Verified, 63.7% SWE-Bench Multilingual, 69.9% Aider—state-of-the-art agentic coding on Together AI
Production-Ready Infrastructure: 256K context with 99.9% SLA on the AI Native Cloud— available on serverless and dedicated endpoints

Quickstart guides

Agents

How to Build Coding Agents

RAG

Building a RAG Workflow

Integrations

Using Together with Vercel's AI SDK

API usage

cURL
Python
Typescript

Endpoint:

Qwen/Qwen3-Coder-Next-FP8

curl -X POST https://api.together.xyz/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -d '{
    "model": "Qwen/Qwen3-Coder-Next-FP8",
    "messages": [{
      "role": "user",
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  }'

from together import Together

client = Together()
response = client.chat.completions.create(
  model="Qwen/Qwen3-Coder-Next-FP8",
  messages=[
  	{
	    "role": "user", 
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }
 ],
)

print(response.choices[0].message.content)

import Together from "together-ai";

const together = new Together();

async function main() {
  const response = await together.chat.completions.create({
    model: "Qwen/Qwen3-Coder-Next-FP8",
    messages: [{
      role: "user",
      content: "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  });
  
  console.log(response.choices[0]?.message?.content);
}

main();

Model card
Architecture Overview:
• Mixture-of-Experts (MoE) architecture with 80B total parameters and 3B activated parameters
• 48 layers with hybrid layout: 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))
• Gated Attention: 16 attention heads for Q, 2 for KV, head dimension 256, rotary position embedding dimension 64
• Gated DeltaNet: 32 linear attention heads for V, 16 for QK, head dimension 128
• 512 experts with 10 activated per token, 1 shared expert, expert intermediate dimension 512
• Hidden dimension 2048 with 79B non-embedding parameters
• 256K context length (262,144 tokens natively)
• Non-thinking mode only—does not generate thinking blocks

Training Methodology:
• Pretraining and post-training stages optimized for coding agents
• Elaborate training recipe for long-horizon reasoning and complex tool usage
• Specialized training for execution failure recovery and dynamic coding tasks
• Trained for seamless integration with diverse development environments

Performance Characteristics:
• Ultra-efficient: 3B activated parameters achieving performance comparable to models with 30-60B active parameters
• Leading agentic coding: 74.2% SWE-Bench Verified (w/ SWE-Agent), 63.7% SWE-Bench Multilingual, 44.3% SWE-Bench Pro
• Strong autonomous coding: 69.9% Aider, 39.3% Terminal-Bench 2.0 (w/ Terminus-2 json)
• Outperforms larger models: beats DeepSeek-V3.2 (37B active), GLM-4.7 (32B active), MiniMax M2.1 (10B active)
• Cost-effective deployment: 10-20x parameter efficiency advantage for agent workloads
• Advanced tool calling capabilities with native support for complex function orchestration
‍
Applications & use cases
Agentic Software Development:
• Production-level autonomous coding: 74.2% SWE-Bench Verified, 63.7% SWE-Bench Multilingual, 44.3% SWE-Bench Pro
• Long-horizon reasoning across complex codebases with 256K context
• Execution failure recovery—adapts when plans don't work as expected
• Multi-step development workflows with precision tool invocation
• Repository-scale navigation and bug fixing
• Code review, refactoring, and optimization tasks

Advanced Tool Calling & Orchestration:
• Native support for complex function calling and tool orchestration
• Dynamic tool selection and sequential execution
• Error handling and recovery from tool execution failures
• Multi-tool workflows for comprehensive development tasks
• Function definition, invocation, and result processing

Autonomous Coding Assistance:
• Code generation from natural language descriptions
• Automated testing and test case generation
• Documentation generation and code commenting
• Debugging and error diagnosis with suggested fixes
• 69.9% Aider performance—strong autonomous coding assistance

Cost-Effective Agent Deployment:
• Ultra-efficient: 3B activated parameters performing like 30-60B models
• 10-20x parameter efficiency advantage reduces infrastructure costs
• Highly cost-effective at $0.50/$1.20 for production agent workloads
• Scales from prototyping to production without cost explosion
• Ideal for startups and enterprises deploying coding agents at scale

Development Workflow Automation:
• End-to-end feature implementation from specification to working code
• Automated code migration and refactoring across files
• Batch processing of code changes across repositories
• CI/CD pipeline integration for automated code generation
• Technical debt reduction through automated refactoring
‍

Related models

Model specifications

Model data

Model provider
Qwen
Type
Code
Chat
LLM
Main use cases
Chat
Coding Agents
Speed
Very High
Intelligence
High
Deployment
Serverless
Monthly Reserved
Endpoint
Qwen/Qwen3-Coder-Next-FP8
Parameters
79.7B
Context length
262K
Input price
$0.50 / 1M tokens
Output price
$1.20 / 1M tokens
Input modalities
Text
Output modalities
Text

Released
February 1, 2026
Last updated
February 2, 2026
Quantization level
FP8
External link
Provider docs
Category
Code

Run in Playground

Quickstart docs

Deploy model

Qwen3-Coder-Next

About model

API usage

Model card

Applications & use cases