阿里发布 Qwen3.8 与 Qwen3.8-Max

2026 年 8 月 3 日，阿里云把 Qwen3.8 与 Qwen3.8-Max 推到台前。总参数 2.4T、每 token 激活约 950B 的稀疏 MoE，原生多模态视觉，1M 上下文——这些数字放在一个月前还像科幻，放在 Kimi K3 之后却只是接力赛的下一棒。7 月 16 日，月之暗面刚用 2.8T 总参数刷新开源规模纪录并登顶 WebDev Arena；不到三周，阿里就带着 2.4T 的旗舰回到牌桌上。

发布材料里最让开发者注意的，不是参数表，而是「下周开源」四个字。官方宣布将开放 Max 权重与 Qwen3.8-27B，截至报道时权重尚未发布——云上旗舰与开放权重被绑进了同一条时间线。过去 Qwen 家族是「开源小模型 + 云上大旗舰」双轨并行，这次阿里第一次承诺把最大档也交出来。对自托管团队来说，这句话的分量比任何榜单都重。

CodeArena 全球第四是官方口径，需要带着厂商自述的边界读。但 16 天的自主开发演示是另一种证据：Qwen3.8 独立完成真实开源项目 oh-my-cli，留下 265 个 commit、127 个 PR、151 个 issue。这不是精心挑选的基准题，而是一个有真实用户、真实 issue 的仓库。Agent 能力的叙事，从「榜单分数」变成了「可核验的工作记录」——虽然演示本身仍由厂商选择与呈现。

同日发布的「千问办公」把模型与产品放在同一场发布会里。API 国内定价每百万 token 输入 12 元、输出 36 元，延续了国产旗舰的性价比路线。模型、应用、价格三件套同时亮相，说明阿里的打法不再是「先发模型再等生态」，而是模型与入口一起交付。

Qwen3.8 的意义不在单个数字，而在节奏本身。Kimi K3 刷新纪录后三周内，阿里就给出回应；开源旗舰的规模竞赛，从「谁先发布」变成了「谁能在对手发布后多久跟上」。对开发者，最强开源模型的选择每隔几周就会变一次；对行业，「开源 + 云」的绑定节奏成了新的竞争变量。下一棒会是谁，已经不需要等太久。

On August 3, 2026, Alibaba Cloud put Qwen3.8 and Qwen3.8-Max on the table. A sparse MoE with 2.4T total parameters and about 950B active per token, native multimodal vision, a 1M-token context—these numbers would have sounded like science fiction a month earlier, but after Kimi K3 they were just the next leg of a relay. On July 16, Moonshot had reset the open-weight scale record at 2.8T and topped WebDev Arena; less than three weeks later, Alibaba was back at the table with a 2.4T flagship.

What developers noticed most was not the parameter table but four words: "open next week." The company said Max weights and Qwen3.8-27B would be released the following week—as of reporting, they had not shipped—which tied the cloud flagship and open weights into one timeline. In the past, the Qwen family ran open small models and cloud big flagships in parallel; this time Alibaba promised to hand over the largest tier as well. For self-hosting teams, that sentence outweighs any leaderboard.

The CodeArena fourth-place rank is an official claim and should be read with that boundary in mind. But the 16-day autonomous demo is a different kind of evidence: Qwen3.8 completed the real open-source project oh-my-cli on its own, leaving 265 commits, 127 PRs, and 151 issues. That is not a curated benchmark—it is a repository with real users and real issues. The agent narrative shifted from leaderboard scores to a verifiable work record, even though the demo itself was vendor-selected and vendor-presented.

The same day, Qianwen Office put model and product in one launch. Domestic API pricing of ¥12 per million input tokens and ¥36 per million output continued the cost-performance line of Chinese flagships. Model, application, and price arriving together shows a playbook that no longer waits for an ecosystem to form: the model and the door are delivered at once.

Qwen3.8 matters not for a single number but for cadence. Within three weeks of Kimi K3's record, Alibaba answered; the open-flagship scale race moved from "who ships first" to "how fast can you follow a rival's launch." For developers, the strongest open model now changes every few weeks; for the industry, the open-plus-cloud cadence became a new competitive variable. The next leg of the relay will not take long to arrive.