[Oseltamivir]

The GLM-5 article equated a cost ratio with a percentage reduction and compared points outside the shared measured range. Recomputing the original historical rows with the dashboard’s unchanged interpolation gives 29.6% lower cost with MTP near 24.3 tok/s/user, and 23.7% without MTP near 11.2 tok/s/user.

Correct both English and Chinese titles, metadata, opening claims, tables and FAQ payloads; identify the actual May 18–19 producer runs behind the May 20 snapshot. Add a dated correction notice and numerical/rendering regressions, and correct the misleading worked example in the blog-authoring guide. This follows up the numerical issue deferred in #1498.

Local smoke and blog E2E runs passed, including the corrected English/Chinese tables at phone and desktop widths. Final Chinese maintainer editorial review remains outstanding.

Data and scope

• Source: GLM-5 benchmark snapshot as of 2026-05-20, filtered to 29 SGLang FP8, single-turn, non-disaggregated, 8192/1024 rows for B200 and MI355X.
• Producers: B200, May 18, and MI355X, May 19. The fixture preserves selected result IDs, source URLs, unrounded metrics and the complete-response hash.
• Historical TCO assumptions remain $1.95/GPU/hr and $1.48/GPU/hr. Build the throughput Pareto frontier, interpolate throughput, derive cost, then compute saving = 1 - MI355X cost / B200 cost before rounding.
• The MTP interpolated maximum is at approximately 24.31953 tok/s/user: cost ratio 1.42092395, reduction 29.6232566%. The table quotes the nearby 24.3 point. The non-MTP maximum is near 11.20183: ratio 1.31080930, reduction 23.7112523%.
• The former 18 tok/s/user MTP and 10 tok/s/user non-MTP comparisons require extrapolation. Replace them with supported targets. Preserve the published slug and image paths so existing links continue to work.
• These are static editorial corrections. Dashboard transformations, API contracts and the inferencex-skills package behavior are unchanged; no public data control is added or altered.

Validation details

• Ran the repository’s local smoke suite with E2E_FIXTURES=1: 60 component checks and 97 integration checks passed, including official/unofficial overlays and Chinese routes.
• Ran cypress/e2e/blog.cy.ts: 17 checks passed, including four new English/Chinese cases at 390px and 1280px that verify rendered costs, metadata, historical links and horizontal overflow.
• FAQ payload parity is checked against the source. The existing MDX renderer does not emit these expression-based FAQ scripts; this PR does not change that rendering behavior. The browser cases verify the emitted BlogPosting metadata.

AI model disclosure

• GPT-6: implementation, benchmark investigation, numerical verification and tests; a delegated GPT-6 agent independently audited the source rows and interpolation extrema. The runtime did not expose verifiable exact model/version identifiers for either agent.
• claude-opus-5-5: advisory review of changed Chinese passages against the revised English and verified numerical results. Its suggested naturalness corrections were applied.

中文

原 GLM-5 文章混淆了成本比值与成本降幅，并比较了超出双方共有实测范围的运行点。使用仪表板现有插值方法重算原始历史数据后，MTP 的最大成本降幅约为 29.6%，出现在约 24.3 tok/s/user；非 MTP 的最大降幅约为 23.7%，出现在约 11.2 tok/s/user。

本次同步修正中英文标题、元数据、开篇结论、表格与 FAQ 数据，标明 5 月 20 日快照所采用的 5 月 18–19 日原始运行，并添加更正说明、数值与渲染回归检查，同时修正博客写作指南中有误的示例。这是对 #1498 中暂缓处理的数值问题的后续修复。

数据保留 29 条源记录的标识符、原始指标和来源链接；TCO 假设仍为 B200 $1.95/GPU/hr、MI355X $1.48/GPU/hr。先对吞吐量前沿插值，再推导成本，最后按 1 - MI355X cost / B200 cost 计算降幅。旧版 MTP 的 18 tok/s/user 与非 MTP 的 10 tok/s/user 比较均需外推，已改为实测范围内的运行点。保留现有文章路径，避免已有链接失效。

本地 smoke 套件的 60 项组件检查和 97 项集成检查通过，博客 E2E 的 17 项检查也全部通过，包含中英文在 390px 和 1280px 下的成本、元数据、历史链接与横向溢出检查。FAQ 数据在源码层验证；现有 MDX 渲染器不会输出这类表达式形式的 FAQ 脚本，本 PR 未改变这一行为。浏览器检查验证实际输出的 BlogPosting 元数据。中文维护者的最终编辑审核尚未完成。

仅修正静态文章和写作指南，不改变仪表板计算、API 契约或 inferencex-skills 行为。GPT-6 负责实现与验证，委派的 GPT-6 agent 独立核对数据和插值极值；运行环境未提供可核实的精确模型版本。claude-opus-5-5 负责中文语义与自然度审核，建议的措辞修正已采纳。

[vercel]

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

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 25, 2026 10:02am 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: 9d161869-c797-4f7a-8437-b4a65d1afec7)