[Oseltamivir]

Chinese technical articles contained mistranslations of compute dies, GPU kernels and host round trips, ambiguous ratio comparisons, and Markdown corruption that hid multiplication signs and damaged SGLANG_OPT_*. This corrects the affected blog, hardware and model copy and restores the README deployment reference. GPU kernel wording is made consistent while the actual OS-kernel module wording is preserved.

The GLM-5 numerical correction is handled separately in #1499 after auditing the original source data: the recomputed maximum saving is about 29.6% with MTP near 24.3 tok/s/user, and 23.7% without MTP near 11.2 tok/s/user. That PR updates both English and Chinese; this PR retains its Chinese-copy-only scope.

Local smoke testing with all three PRs combined passed: 60 component checks and 129 integration checks, including Chinese desktop/mobile paths and official/unofficial-run overlays. Server-rendered checks also confirmed the repaired multiplication expression, SGLANG_OPT_*, and API fallback descriptions.

Validation details

Ran the repository’s quick component and integration suites plus the fleet lifecycle E2E spec against the combined changes with E2E_FIXTURES=1. The component assertions were updated to follow the corrected 投机解码 copy. HTTP checks returned 200 and verified rendered content on /zh/blog/inferencemax-open-source-inference-benchmarking, /zh/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days, and /zh/api. This is automated coverage; it does not replace the outstanding human editorial and visual review.

Review scope

The changes affect Chinese prose, metadata, image descriptions and FAQ answers. English content and benchmark data are unchanged. API parameters, returned values and calculations are unchanged; this is presentation-only copy, so no API or skills contract change is needed. The added regression checks compile and render the affected Markdown; both fail against the original broken paragraphs.

Claude reviewed the changed passages against their English sources using .claude/skills/review-zh-copy/SKILL.md. Its findings were addressed or deferred with the source inconsistency above. Desktop and mobile visual review and final Chinese maintainer editorial review have not been completed.

AI model disclosure

GPT-6 prepared and verified the changes; the runtime does not expose a more exact model/version identifier, so it could not be verified. Claude Opus 5.5 (claude-opus-5-5, verified from CLI model usage metadata) reviewed Chinese fidelity and naturalness. A contributing GPT-6 parent agent coordinates this PR with the separate dashboard and technical-reference fixes; its exact version could not be verified either.

中文

本地将三个 PR 的改动合并后运行 smoke 检查，60 项组件检查和 129 项集成检查全部通过，覆盖中文桌面端、移动端及官方/非官方运行叠加。服务端渲染检查也确认了乘法表达式、SGLANG_OPT_* 和 API 回退说明。自动化检查不能替代尚未完成的人工编辑与视觉审核。

中文技术文章中存在 compute die、GPU kernel 和主机往返传输等概念误译，部分倍数表达有歧义，Markdown 还破坏了乘号和 SGLANG_OPT_* 标识符。本次修正相关博客、硬件与模型页面文案，补回 README 的部署配置文档链接，并统一 GPU kernel 用语，同时保留真正的操作系统内核表述。

GLM-5 数值问题已在核对原始数据后交由 #1499 单独修正：重算后的最大成本降幅为 MTP 下约 24.3 tok/s/user 处的 29.6%，以及非 MTP 下约 11.2 tok/s/user 处的 23.7%。该 PR 同步更新中英文，本 PR 继续仅修改中文文案。

仅修改中文展示文案；英文内容、基准数据、API 参数、返回值和计算逻辑均保持不变。新增回归测试实际编译并渲染损坏过的 Markdown，两个测试在原始内容上均失败。Claude 已按仓库 skill 审核语义与自然度；桌面端、移动端视觉检查及中文维护者的最终编辑审核尚未完成。AI 使用情况见上方披露。

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 25, 2026 9:43am UTC

[cursor]

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: ef7bd7f9-be4c-4c4d-a3a5-c07d0e011f0a)