qwen3.8-max

Qwen3.8-Max was released on August 3, 2026. It is the largest model in the Qwen series to date, with a total parameter volume of 2.4 trillion. It adopts a sparse MoE architecture and only activates about 95 billion parameters during inference, achieving a balance between ultra-large scale and inference efficiency. The model supports 1 million Token context windows, ranks fifth in Text Arena and second in Vision Arena, and has native visual intelligence and multi-modal understanding capabilities. Qwen3.8-Max has demonstrated cutting-edge capabilities in autonomous programming and long-term tasks. During internal testing, it independently completed a 16-day software engineering project and built and open sourced the self-evolving agent framework "oh-my-cli" from scratch. The model opens API calls to global developers through the Alibaba Cloud Bailian platform, and the weights are planned to be open source.

Context
1000000 tokens
Max output
131072 tokens
Input
llm image video
Output
llm

Pricing

Input: $2 / 1M tokens

Output: $6 / 1M tokens

Cached input: $0.25 / 1M tokens

Cache write: $2.5 / 1M tokens

Endpoints

  • openai

More models

  • gemini-3.5-flash-lite

    gemini-3.5-flash-lite 是 Gemini 3.5 Flash 系列中的轻量模型，面向需要以较低资源成本完成大量重复性请求的场景。它适合做内容初筛、短摘要、字段规范化、语种识别、文本路由、模板回复和候选结果排序。将其置于多模型流程的前端，能够先过滤简单请求，再把疑难或高价值任务交给能力更强的模型。上线前应以真实流量样本检验误分率、输出一致性和异常输入表现，并保留人工抽检机制。

  • gpt-4.1

    gpt-4.1 是 OpenAI GPT-4.1 系列模型标识。官方目录将 GPT-4.1 定位为非推理型通用模型，适用于文本处理、指令遵从、代码辅助和需要稳定、直接响应的业务场景。它适合客服知识问答、结构化信息提取、内容改写、文档摘要和工具调用前后的语言处理。设计应用时应使用清晰的输出 schema、代表性示例和错误兜底；对于需要多步深度推理的任务，应比较更适合推理的模型。

  • claude-opus-5-5
  • qwen3.6-plus

    Qwen3.6-Plus 于 2026 年 4 月 2 日发布，是千问 3.6 系列的首款模型，被官方称为当时“中国编程能力最强的模型”。它在 Qwen3.5 基础上实现了智能体编程能力的显著跃升，在 SWE-bench Verified 上得分 78.8，Terminal-Bench 2.0 得分 61.6，表现超越参数量为其两到三倍的 GLM-5、Kimi-K2.5 等国产模型，接近 Claude 系列的水平。模型默认支持 100 万 Token 上下文窗口，深度融合了推理、记忆与执行能力，在代码修复、终端自动化和长程规划任务中展现出稳定的工程落地能力。该版本通过阿里云百炼 API 开放调用，定位为面向开发者的“氛围编程”基础设施。