Qwen4 Max
Flagship Qwen model for long-running agentic work with 1M context

About model
Qwen4 Max is Alibaba's next-generation flagship Qwen model, which Alibaba describes as the most capable model in the Qwen family to date. It is a sparse Mixture-of-Experts model with 2.4 trillion total parameters, activating 71 billion per token, which Alibaba reports is 25% fewer than Qwen3.8 Max at the same scale. It is built on the new Qwen4 architecture, which upgrades the design across attention, residual connections, and embeddings. The model takes text and image input, has deep thinking enabled by default, and supports a 1M-token context window. Alibaba positions it as an AI teammate for long-running agentic work across software engineering, professional fields, creative production, and scientific research. Available on Together AI.
2.4T
Sparse MoE on the new Qwen4 architecture
1M
Long agent sessions, large codebases, and document sets in one request
Native
Images processed alongside text, with deep thinking on by default
- Long-Running Agentic Work: Positioned for multi-step tasks across software engineering, finance, law, healthcare, and scientific research
- Secure Software Engineering: Trained with code security checks alongside functionality tests, targeting vulnerabilities such as injection and broken access control
- Creative & Technical Production: Supports 3D modeling, game development, and agentic video creation alongside image understanding
- Available on Together AI: Serverless access through the Together AI API
API usage
Endpoint:
Model card
Architecture Overview:
• Sparse Mixture-of-Experts model with 2.4T total parameters and 71B activated per token
• Built on the Qwen4 architecture, which upgrades the Mixture-of-Experts design across attention, residual connections, and embeddings
• Native vision with text output, across a 1M-token context window
Training Methodology:
• Software engineering training paired functionality tests with code security checks to reduce vulnerabilities in generated code
• Cybersecurity data filtered by task intent, keeping defensive material and removing offensive attack content
• Purpose-built safety datasets and post-training enhancements addressing chemical, biological, radiological, and nuclear risks
Performance Characteristics:
• Alibaba reports that on Work360, its in-house benchmark spanning 721 specialized fields, Qwen4 Max outperforms Qwen3.8 Max overall while using 35% fewer tokens in high-effort reasoning mode and cutting average cost per task by about 60%
• Alibaba-reported demonstrations include a 48-hour autonomous agent workflow that produced an experimentally validated coolant molecule, and a manufacturing case that reduced manual workpiece programming time by 95%
• Benchmark tables arrive with Alibaba's full technical materials
Prompting
Together AI API Access:
• Access Qwen4 Max via Together AI APIs using the endpoint Qwen/Qwen4-Max
• Authenticate using your Together AI API key in request headers
• Deep thinking is enabled by default
• Supports text and image input across a 1M-token context window
• Available on Together AI with serverless access
Applications & use cases
Agentic Software Engineering:
• Run long-horizon coding agents across large repositories
• Generate code with security in mind from the start
• Hold full projects in the 1M-token window
Professional & Industry Work:
• Support finance, legal, and healthcare workflows end to end
• Plan production lines and verify part designs in manufacturing
• Automate multi-step operational tasks
Research & Creative Production:
• Run long-running scientific research agents with many tool calls
• Build 3D models, games, and video projects with agentic workflows
• Reason over images and documents alongside text
- TypeChatReasoningVision
- Deployment
- Endpoint
- Parameters2.4T
- Activated parameters71B
- Context length1M
- Input price
$1.43 / 1M tokens
$0.0713 (cached)/1M
- Output price
$4.28 / 1M tokens
- Input modalitiesTextImage
- Output modalitiesText
- ReleasedNovember 15, 2026
- CategoryChat
