fix(models): correct four release dates and source the other eight / 修正四个错误的模型发布日期并为其余八个补充来源

[Oseltamivir]

#738 and #742 have merged. Rebased directly onto master — the squash-merge of #738 landed as a new commit, so this branch was still carrying the original and GitHub saw the same change twice. Now a single commit on master, no conflicts.

Summary

#738 moved the release dates into one table without changing them. This PR fixes them. Four of the eight were wrong, and four models had no date at all.

Every entry now cites its source and records the model's first InferenceX sweep (MODELS.md "Date added") beside it, so the invariant that caught these is checkable at review time: a model cannot be benchmarked before its weights exist.

Corrected

Model	Was	Now	Why the old value is wrong
DeepSeek-V4-Pro	2026-06-08	2026-04-24	Six weeks after its own first sweep (2026-04-25). V4-Pro-0813 went GA 2026-08-13, but Hugging Face still hosts the April preview build, so the weights date does not move.
Kimi-K3	2026-06-13	2026-07-27	Weights day. The product launched 2026-07-16; the gap let vLLM/NVIDIA/AMD prepare day-zero support, and InferenceX swept it the day weights landed.
MiniMax-M2.5	2025-10-25	2026-02-12	That is M2's launch, not M2.5's. M2.5 is architecturally unchanged from M2, and the architecture entry still points its sourceUrl at MiniMax-M2 — which is where the date came from.
gpt-oss-120b	2025-06-13	2025-08-05	The Apache 2.0 release alongside gpt-oss-20b. The old date precedes it by seven weeks and matches no published event.

Added

Qwen-3.5-397B-A17B 2026-02-16 · GLM-5 2026-02-13 · MiniMax-M3 2026-06-07 · GLM-5.2 2026-06-13

Unchanged, confirmed

DeepSeek-R1-0528 2025-05-28 (the date the name encodes) · Kimi-K2.5 2026-01-27 · Llama-3.3-70B 2024-12-06 · Llama-3.1-70B 2024-07-23

Two dates are softer than the rest

Both say so in place rather than presenting a guess as sourced:

• GLM-5 — weights sourced only to "mid-February"; reporting splits between 2026-02-13 and 2026-02-17. The HF commit history at zai-org/GLM-5 is the authority if it ever needs to be exact.
• GLM-5.2 — most sources date the HF upload 2026-06-13, one says 2026-06-16.

Each is ±3 days on an axis sampled monthly, and both sit comfortably before their first sweep, so the ambiguity changes no rendered value — but these are announcement-day dates, not confirmed upload days.

Conventions applied

• The date is the weights publication, not the announcement and not the day InferenceX started benchmarking. Third-party inference optimisation — exactly what the Fleet Lifecycle chart tracks — cannot begin before the weights are public.
• A point-release bucket takes the earliest release in it, because that is when the bucket's weights first existed: GLM-5 covers 5 and 5.1, MiniMax-M2.5 covers 2.5 and 2.7, Kimi-K2.5 covers 2.5, 2.6 and 2.7-Code.

Why it matters

Every selectable model now has a date, so the Fleet Lifecycle axis stops falling back to "anchored to its first benchmark run instead" — which reads as when the model appeared but means when we got to it. For gpt-oss that fallback was five weeks late; for Llama-3.3, eight months.

On master the visible change is the architecture-diagram caption for the four corrected models, plus MiniMax-M3 gaining a caption it never rendered.

Verification

Both new tests were mutation-verified — mutant applied, named test observed failing, reverted:

Test	Mutation that kills it
coverage over MODEL_OPTIONS	delete the Qwen-3.5-397B-A17B entry
the twelve dates pinned	revert gpt-oss to 2025-06-13

• packages/constants unit: 39/39 · model-architectures.test.ts: 68/68
• bun run lint, bun run fmt, bun run typecheck: clean
• Pre-existing failures on master, untouched here: api-route-catalog digest, overview-data AgentX frontier, useThroughputData agentic interpolation.

中文说明

#738 只是把发布日期合并到一张表，未改动数值；本 PR 负责修正。八个日期中四个是错的，另有四个模型完全没有日期。现每条均附来源，并在旁记录 MODELS.md 中该模型的首次 InferenceX 基准日期，以便评审时核对「模型不可能早于其权重发布被基准测试」这一约束。

修正：DeepSeek-V4-Pro 2026-06-08 → 2026-04-24（原值比首次基准 2026-04-25 还晚六周；0813 版本虽已 GA，但 Hugging Face 仍托管四月预览版权重，故日期不变）；Kimi-K3 2026-06-13 → 2026-07-27（权重发布日，产品发布为 07-16，留出时间供 vLLM/NVIDIA/AMD 准备零日支持）；MiniMax-M2.5 2025-10-25 → 2026-02-12（原值是 M2 的发布日而非 M2.5；M2.5 架构与 M2 相同，架构条目的 sourceUrl 至今仍指向 MiniMax-M2，错误正源于此）；gpt-oss-120b 2025-06-13 → 2025-08-05（Apache 2.0 发布日，原值早了七周且不对应任何公开事件）。

新增：Qwen-3.5-397B-A17B 2026-02-16、GLM-5 2026-02-13、MiniMax-M3 2026-06-07、GLM-5.2 2026-06-13。确认无误：DeepSeek-R1-0528、Kimi-K2.5、Llama-3.3-70B、Llama-3.1-70B。

其中 GLM-5 与 GLM-5.2 的权重上传日仅能定位到 ±3 天（GLM-5 仅有"二月中"的说法，报道分散在 02-13 与 02-17；GLM-5.2 多数来源为 06-13，一处为 06-16），已在注释中如实标明，不将推测包装为确证。

约定：日期取权重公开可下载之日，而非发布公告日或 InferenceX 开始基准测试之日——第三方推理优化（也正是集群生命周期图所追踪的内容）只能在权重公开后才可能开始；分桶取其中最早的发布日期。

现所有可选模型均有日期，集群生命周期时间轴不再回退到"以首次基准测试锚定"——该回退会被读作"模型出现的时间"，实则是"我们开始测的时间"，gpt-oss 相差五周，Llama-3.3 相差八个月。在 master 上可见的变化是四个模型的架构图日期说明，以及 MiniMax-M3 首次显示该说明。

验证：两项新测试均经变异测试验证；constants 单测 39/39，model-architectures.test.ts 68/68，lint/fmt/typecheck 均通过。另有三项单测在 master 上即为失败，与本 PR 无关。

---

Note

Low Risk
Internal constants and UI date anchoring only; no API contract or auth changes. Wrong dates would mislabel charts but are now heavily tested.

Overview
Corrects and completes MODEL_RELEASE_DATES so Fleet Lifecycle and architecture captions anchor on weights publication, not first InferenceX sweep or wrong sibling launches.

Four dates are fixed: DeepSeek-V4-Pro (2026-04-24, before its first sweep), Kimi-K3 (2026-07-27), MiniMax-M2.5 (2026-02-12, not M2’s date), and gpt-oss-120b (2025-08-05). Four models gain sourced entries (Qwen-3.5, GLM-5, MiniMax-M3, GLM-5.2), each with inline sources and sweep: notes for review.

Tests now require every MODEL_OPTIONS model to have a release date, pin twelve expected dates (mutation-tested), and expect full architecture-caption coverage (e.g. MiniMax-M3). Constants tests document the null fallback for unknown names only.

API contract ledger updates the models.ts digest and review note—release-date values only; no published API surface change.

^{Reviewed by Cursor Bugbot for commit 66721e0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 15, 2026 6:13am

[Oseltamivir]

Rebased so the stack is #742 → #738 → this. Base retargeted to models/single-release-date-table.

Same digest note as #738: this PR edits ../constants/src/models.ts again (date values inside MODEL_RELEASE_DATES only), so its API-doc digest is updated a second time. Reviewed — no published model name, alias, or parameter enum changes, and no endpoint exposes a release date.

Unit suite at this commit: 3368 app + 39 constants, all passing.

中文：已重新 rebase，堆叠顺序为 #742 → #738 → 本 PR，base 改为 models/single-release-date-table。与 #738 相同：本 PR 再次修改 ../constants/src/models.ts（仅 MODEL_RELEASE_DATES 内的日期值），故其 API 文档摘要需再次更新。经复核：未改动任何已发布的模型名称、别名或参数枚举，也没有任何端点暴露发布日期。当前提交单测：app 3368 项 + constants 39 项，全部通过。