Qwen3.8-Max
Overview
Qwen3.8-Max is Alibaba's flagship Qwen model for coding, professional knowledge work, multimodal understanding, and long-horizon agent tasks. It was officially released on August 3, 2026, with the API model ID qwen3.8-max. The model uses a Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters and 95 billion activated parameters, accepts text, image, and video input, and produces text output.
Key Features
- 2.4T-Parameter MoE Architecture: Built on the Qwen3.5 architectural foundation with 2.4 trillion total parameters and 95 billion activated parameters per token.
- Native Multimodal Intelligence: Processes text, images, and video in the same model. Qwen describes vision as part of an agent's planning, execution, and self-correction loop rather than a separate preprocessing step.
- 1M-Token Context: Provides a 1-million-token context window, with up to 991.80K input tokens in non-thinking mode, 983.61K input tokens in thinking mode, and 131.07K output tokens.
- Configurable Reasoning: Thinking is enabled by default. The API supports
low,medium, andxhighreasoning effort, withxhighas the default, and preserves reasoning across turns by default. - Long-Horizon Agent Work: The launch report documents autonomous coding, research, chip-design optimization, and year-long business-simulation tasks. In one coding case, the model operated for about 16 days and produced 265 commits, 127 pull requests, and 151 issues without human intervention.
- Agent and API Integration: Supports Function Calling, structured output, prefix completion, batch inference, context caching, and built-in tools. It can be accessed through OpenAI-compatible Chat Completions and Responses APIs, Anthropic-compatible APIs, or DashScope.
Best Use Cases
- Autonomous Software Engineering: Multi-day implementation, repository maintenance, testing, debugging, and iterative delivery using coding agents.
- Research and Experimentation: Reproducing papers, writing experimental pipelines, running iterative evaluations, and refining hypotheses with tools.
- Professional Knowledge Work: Legal review, financial analysis, document production, data analysis, design prototyping, and other multi-step office workflows.
- Long-Horizon Planning and Optimization: Tasks that require hundreds of feedback cycles, persistent state, adaptive planning, and repeated verification.
- Multimodal Agents: Workflows that combine documents, screenshots, images, video, code execution, search, and external tools.
Capabilities and Limitations
| Capability | Description |
|---|---|
| Reasoning | Thinking is enabled by default. reasoning_effort supports low, medium, and xhigh; xhigh is the default. The maximum mapped thinking budget is 262,144 tokens, while the default budget is 131,072. |
| Coding | Designed for autonomous, long-horizon coding. Qwen reports 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 73.5 on FrontierSWE, and 93.0 on PaperBench. |
| Multimodal | Accepts text, image, and video input and produces text output. |
| Context Window | 1M tokens. |
| Maximum Input | 991.80K tokens in non-thinking mode and 983.61K tokens in thinking mode. |
| Max Output | 131.07K tokens in both modes. |
| Structured Output | Supported. |
| Architecture | Mixture-of-Experts architecture based on Qwen3.5. |
| Total Parameters | 2.4 trillion. |
| Activated Parameters | 95 billion. |
| Knowledge Cutoff | Unpublished. |
Known Limitations
- Preserved thinking is enabled by default. Multi-turn clients must return the complete, unmodified
reasoning_contenthistory; preserved reasoning also counts toward input tokens and billing. - Alibaba has not published a model-specific knowledge cutoff, measured generation speed, or complete language-coverage list.
Standard Pricing
The token prices below are shown in USD per 1 million tokens; web search is billed in USD per use.
| Model | Input (USD / 1M Tokens) | Cache Write (USD / 1M Tokens) | Cache Read (USD / 1M Tokens) | Output (USD / 1M Tokens) | Web Search (USD / use) | Billing Notes |
|---|---|---|---|---|---|---|
| Qwen3.8-Max | $2.00 | $2.00 | $0.25 | $6.00 | - | Cache Read uses the official implicit-cache price. |
Credits settlement: B.AI converts charges at 1 USD = 1,000,000 Credits and deducts Credits from the account balance.
Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through account benefits. Specific prices and account benefits are subject to the platform display and final billing records.