给大模型一句指令"创建一个自进化的智能体Harness"，然后等16天——回来的时候，它已经从一个空文件夹里写出了一个完整的、可生产使用的Agent框架。这不是科幻，这是阿里今天发布的Qwen3.8-Max干出来的事。
8月3日，阿里巴巴正式发布新一代旗舰基座大模型Qwen3.8-Max，总参数量2.4万亿（MoE架构，激活95B），上下文长度1M Tokens，支持视觉理解。Arena全球排行榜上，Qwen系列仅次于Claude，位列全球第一梯队。但真正炸场的不是参数规模，是它的自主编程能力。
16天：从空文件夹到开源Agent框架
AI编程工具进化到了什么程度？两年前的前沿模型能帮你补全函数、写个小组件；去年的模型能根据需求写出完整的应用代码；Qwen3.8-Max做的事情是：接到一句指令，自己搭建循环工程框架，编排智能体自动领取任务、执行任务、检查结果、修复bug，在16天几乎无人干预的情况下，从零完成了一个Hermes Agent级别的真实可用框架——oh-my-cli。
项目已完全开源，完整过程公开在GitHub上。人类工程师可以通过钉钉给Qwen3.8提新要求，但核心开发工作是模型自主完成的。这意味着AI编程的范式已经从"AI辅助写代码"进化到了"AI自主交付项目"——你不需要逐行review代码，你只需要验收最终产物。
评测数据支撑了这个能力跃迁：PaperBench科研复现评测中，Qwen3.8-Max拿到93.0分，较上代提升28.2分，创下评测新高；CodeArena全球第四；评估智能体电脑操作能力的OSWorld-Verified以86.1分位居主流模型首位；指令遵循IF Bench 82.8分，科学推理GPQA Diamond 92.6分。
定价直接砍到Opus5的四分之一
如果能力是弹药，定价就是扣扳机的手指。Qwen3.8-Max的API定价策略攻击性十足：国内输入12元/百万Tokens，输出36元/百万Tokens，隐式缓存命中低至1.5元/百万Tokens。国际价格更狠：输入为Claude Opus5的40%，输出仅为24%——不到四分之一的价格，提供第一梯队的能力。
同时发布的还有Agent产品"千问办公"，定位专业办公场景。Qwen3.8-Max的视觉能力也大幅提升，Vision Arena排名全球第二，不仅能读懂200页PDF财报、看懂超100小时长视频，还能在无源代码、无联网的情况下，仅凭交互反馈从零复刻整个应用。
开源计划同样激进：Qwen3.8-Max预计下周开源，同时开源Qwen3.8-27B。阿里真武M890超节点已完成适配，Agentic推理场景最多可实现1.5倍性能提升。
从GPT-4时代的"AI帮你写代码"到今天的"AI自己交付项目"，这个进化速度比大多数人预期的快得多。当一个模型可以连续工作16天、自主搭建框架、修复bug、交付可用产品的时候，程序员的角色确实在从"写代码的人"变成"给AI提需求、验成果的人"。
明天见。
Give a large model one instruction — "create a self-evolving agent harness" — then wait 16 days. When you come back, it has built a complete, production-ready agent framework starting from an empty folder. This isn't science fiction. It's what Alibaba's Qwen3.8-Max, released today, actually did.
On August 3, Alibaba officially launched its new flagship base model Qwen3.8-Max: 2.4 trillion total parameters (MoE architecture, 95B activated), 1M token context window, vision understanding support. On the Arena global leaderboard, the Qwen series trails only Claude, firmly in the global top tier. But the bombshell isn't parameter count — it's autonomous coding capability.
16 Days: Empty Folder to Open-Source Agent Framework
How far have AI coding tools come? Two years ago, frontier models autocomplete functions and write widgets. Last year's models could produce full application code from specs. What Qwen3.8-Max does: given one instruction, it sets up a loop engineering framework on its own, orchestrates agents to claim tasks, execute them, check results, fix bugs, and over 16 days with essentially no human intervention, builds a production-grade framework at Hermes Agent level from scratch — oh-my-cli.
The project is fully open source on GitHub with the complete development history preserved. Human engineers can send new requirements to Qwen3.8 via DingTalk, but the core development was done autonomously. This means AI coding's paradigm has shifted from "AI-assisted coding" to "AI-autonomous project delivery" — you don't review code line by line; you accept the final deliverable.
Benchmarks back this capability leap: on PaperBench research reproduction, Qwen3.8-Max scored 93.0, up 28.2 points from the previous generation — a new record. CodeArena: #4 globally. OSWorld-Verified (agent computer operation): 86.1, the top among mainstream models. IF Bench instruction following: 82.8. GPQA Diamond scientific reasoning: 92.6.
Pricing Slashed to a Quarter of Opus 5
If capability is the ammunition, pricing is the finger on the trigger. Qwen3.8-Max's API pricing is aggressively competitive: China domestic input at 12 yuan/MTok, output at 36 yuan/MTok, implicit cache hits as low as 1.5 yuan/MTok. International pricing is even more aggressive: 40% of Claude Opus 5 for input, just 24% for output — top-tier capability at under a quarter the price.
Also launching is the agent product "Qwen Cowork," targeting professional office scenarios. Vision capabilities also improved dramatically — Vision Arena #2 globally, capable of reading 200-page financial PDFs, understanding 100+ hour videos, and even replicating entire applications from scratch through interactive feedback alone — no source code, no internet access required.
The open-source roadmap is equally aggressive: Qwen3.8-Max is expected to be open-sourced next week, alongside Qwen3.8-27B. Alibaba's Zhenwu M890 super-node has already been adapted, delivering up to 1.5x performance gains in Agentic inference scenarios.
From the GPT-4 era of "AI helps you code" to today's "AI delivers projects autonomously," the pace is faster than most expected. When a model can work continuously for 16 days, set up its own framework, fix bugs, and ship usable products — the programmer's role is genuinely shifting from "code writer" to "specifier and acceptor."
See you tomorrow.
AI编程从'帮你写代码'到'自己交付项目'——16天无人干预从零写出一个Agent框架，这是范式切换。
—— Dawn Vision编辑部
AI coding moves from 'helping you write code' to 'delivering projects on its own' — 16 days of building an agent framework from zero with no human intervention is a paradigm shift.
— The Dawn Vision Editorial Desk
Qwen3.8-Max · 阿里千问 · 2.4万亿参数 · 自主编程16天 · oh-my-cli · PaperBench 93分 · AI编程范式迁移 · 千问办公 · 模型定价战
Qwen3.8-Max · Alibaba Qwen · 2.4T parameters · 16 days autonomous coding · oh-my-cli · PaperBench 93 · AI coding paradigm shift · Qwen Cowork · model pricing war
Sources · 信源 Sources
- 潮新闻 - 阿里"超大杯"Qwen3.8正式发布
- IT之家 - Qwen3.8-Max正式上线
- GitHub - oh-my-cli开源仓库
本文基于 Dawn Vision 认知引擎处理的 12 个源信号生成，经编辑部人工审核。素材来源：潮新闻、IT之家、GitHub。
This article was generated by the Dawn Vision cognitive engine processing 12 source signals, with human editorial review. Sources: Chao News, IT Home, GitHub.
