fix(models): correct four release dates and source the other eight

Four of the eight dates carried over from `model-architectures.ts` were wrong,
and four models had none at all. Each entry now cites its source, and records
the model's first InferenceX sweep from MODELS.md beside it so the invariant that
caught these — a model cannot be benchmarked before its weights exist — is
checkable at review time.

Corrected:

- DeepSeek-V4-Pro 2026-06-08 → 2026-04-24. The old date fell six weeks after the
  first sweep on 2026-04-25. V4-Pro-0813 went GA on 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, when the weights were published. The product
  launched 2026-07-16; the eleven-day gap was to let vLLM, NVIDIA and AMD prepare
  day-zero support, and InferenceX swept it the day the weights landed.
- MiniMax-M2.5 2025-10-25 → 2026-02-12. The old date 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 and confirmed: DeepSeek-R1-0528, Kimi-K2.5,
Llama-3.3-70B, Llama-3.1-70B.

Two entries are softer than the rest and say so in place. GLM-5's weights are
sourced only to "mid-February", with reporting split between 2026-02-13 and
2026-02-17; GLM-5.2's Hugging Face upload is dated 2026-06-13 by most sources and
2026-06-16 by one. Both are ±3 days on an axis sampled monthly, and both are
comfortably before their first sweep, so the ambiguity changes nothing — but it
is an announcement-day date, not a confirmed upload day.

A bucket that groups point releases takes the earliest release in it, since 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.

Every selectable model now has a date, so the Fleet Lifecycle axis stops falling
back to "anchored to its first benchmark run" — for gpt-oss that fallback was
five weeks late. Two tests, both mutation-verified: coverage over MODEL_OPTIONS
(dropping Qwen-3.5 fails it), and the twelve dates pinned (reverting gpt-oss to
2025-06-13 fails it).

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

中文：从 `model-architectures.ts` 迁移过来的八个日期中有四个是错的，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，
便于评审时核对「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（原值比首次基准还晚六周）；Kimi-K3 2026-06-13 → 2026-07-27
（权重发布日，产品发布为 07-16）；MiniMax-M2.5 2025-10-25 → 2026-02-12（原值是 M2 的
发布日，非 M2.5）；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、GLM-5、MiniMax-M3、GLM-5.2 四个日期。
GLM-5 与 GLM-5.2 的权重上传日仅能定位到 ±3 天，已在注释中如实标明。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首次基准日期。两项测试
均经变异测试验证。

Author: Oseltamivir

packages/app/src/lib/model-architectures.test.ts

@@ -1,7 +1,7 @@
 import { MODEL_RELEASE_DATES, getModelReleaseDate } from '@semianalysisai/inferencex-constants';
 import { describe, expect, it } from 'vitest';
 
-import { Model } from '@/lib/data-mappings';
+import { MODEL_OPTIONS, Model } from '@/lib/data-mappings';
 
 import {
   type SubBlockFlow,
@@ -21,7 +21,35 @@ import {
   MODEL_ARCHITECTURES,
 } from './model-architectures';
 
-describe('MODEL_RELEASE_DATES keys', () => {
+describe('MODEL_RELEASE_DATES coverage', () => {
+  it('has a sourced date for every model the user can select', () => {
+    // Without a date the Fleet Lifecycle axis falls back to the model's first
+    // benchmark run, which reads as "this is when the model appeared" when it is
+    // really "this is when we got to it" — for gpt-oss that gap is five weeks.
+    // A model whose release date genuinely cannot be sourced belongs in an
+    // explicit allowlist here, so the gap stays visible; do not delete the test.
+    const missing = MODEL_OPTIONS.filter((model) => !getModelReleaseDate(model));
+    expect(missing).toEqual([]);
+  });
+
+  it('anchors each model to its weights publication, not its first sweep', () => {
+    // Pinned so a silent revert fails. Every one of these predates the model's
+    // first InferenceX sweep, which is the invariant that caught the four wrong
+    // dates: a model cannot be benchmarked before its weights exist.
+    expect(getModelReleaseDate(Model.Kimi_K3)).toBe('2026-07-27');
+    expect(getModelReleaseDate(Model.GLM_5_2)).toBe('2026-06-13');
+    expect(getModelReleaseDate(Model.MiniMax_M3)).toBe('2026-06-07');
+    expect(getModelReleaseDate(Model.DeepSeek_V4_Pro)).toBe('2026-04-24');
+    expect(getModelReleaseDate(Model.GLM_5)).toBe('2026-02-13');
+    expect(getModelReleaseDate(Model.Qwen3_5)).toBe('2026-02-16');
+    expect(getModelReleaseDate(Model.MiniMax_M2_5)).toBe('2026-02-12');
+    expect(getModelReleaseDate(Model.Kimi_K2_5)).toBe('2026-01-27');
+    expect(getModelReleaseDate(Model.GptOss)).toBe('2025-08-05');
+    expect(getModelReleaseDate(Model.DeepSeek_R1)).toBe('2025-05-28');
+    expect(getModelReleaseDate(Model.Llama3_3_70B)).toBe('2024-12-06');
+    expect(getModelReleaseDate(Model.Llama3_1_70B)).toBe('2024-07-23');
+  });
+
   it('names a real model, so no entry is dead weight', () => {
     // The table is keyed by string because it lives in the constants package,
     // below the `Model` enum. A key that matches no model never throws — the
@@ -72,12 +100,11 @@ describe('MODEL_ARCHITECTURES', () => {
 
   it('sources every architecture caption from the one release-date table', () => {
     // The diagram prints "Released by {developer} on {date}", so an entry with a
-    // developer and no date silently drops the whole caption. Listing the models
-    // that have no sourced date keeps that visible instead of invisible.
+    // developer and no date silently drops the whole caption.
     const missing = Object.values(MODEL_ARCHITECTURES)
       .filter((arch) => arch?.developer && !getModelReleaseDate(arch.model))
       .map((arch) => arch!.model);
-    expect(missing).toEqual([Model.MiniMax_M3]);
+    expect(missing).toEqual([]);
   });
 
   it('ensures dense models have equal active and total params', () => {

packages/constants/src/models.test.ts

@@ -70,8 +70,10 @@ describe('MODEL_RELEASE_DATES', () => {
   });
 
   it('returns null rather than throwing for a model with no sourced date', () => {
-    expect(getModelReleaseDate('GLM-5.2')).toBeNull();
+    // Every model shipped today has a date, so this documents the fallback path
+    // that a future model arrives on rather than an existing gap.
     expect(getModelReleaseDate('not-a-model')).toBeNull();
+    expect(getModelReleaseDate('')).toBeNull();
   });
 });
 

packages/constants/src/models.ts

@@ -105,14 +105,66 @@ export function rowToSequence(row: {
  * earliest date they actually have data for.
  */
 export const MODEL_RELEASE_DATES: Record<string, string> = {
+  // Newest first. `sweep:` is the model's "Date added" from MODELS.md — the first
+  // InferenceX benchmark — recorded so the "cannot predate its own weights"
+  // invariant above is checkable by eye at review time.
+  //
+  // Weights published on Hugging Face 2026-07-27 under the Kimi K3 License,
+  // eleven days after the 2026-07-16 product launch; the gap was to let vLLM,
+  // NVIDIA and AMD prepare day-zero support. sweep: 2026-07-27 — day zero.
+  'Kimi-K3': '2026-07-27',
+  // Zhipu launched GLM-5.2 on 2026-06-13 via the GLM Coding Plan with MIT
+  // weights at zai-org/GLM-5.2; standalone API and provider support followed
+  // over the next few days. One source dates the Hugging Face upload 2026-06-16
+  // instead, so this is ±3 days — immaterial on a multi-month axis, but it is an
+  // announcement-day date rather than a confirmed upload day. sweep: 2026-07-18.
+  'GLM-5.2': '2026-06-13',
+  // Weights and the official MXFP8 quant live on Hugging Face at
+  // MiniMaxAI/MiniMax-M3 by 2026-06-07, under the custom minimax-community
+  // license. The 2026-06-01 launch shipped API access only; the arXiv report
+  // followed on 2026-06-11. sweep: 2026-06-12.
+  'MiniMax-M3': '2026-06-07',
+  // V4 Preview — V4-Pro (1.6T total / 49B active) and V4-Flash (284B/13B)
+  // published together under MIT on Hugging Face, the API and chat.deepseek.com.
+  // V4-Pro-0813 went GA on 2026-08-13, but the Hugging Face repo still hosts the
+  // April preview build, so the weights date is unchanged. sweep: 2026-04-25.
+  'DeepSeek-V4-Pro': '2026-04-24',
+  // Bucket covers GLM-5 and GLM-5.1, so the date is GLM-5's: released
+  // 2026-02-13, 744B/40B active, MIT, trained entirely on Huawei Ascend. Sources
+  // put the weights in "mid-February" without a day, and reporting clusters on
+  // 2026-02-13 and 2026-02-17; the Hugging Face commit history at zai-org/GLM-5
+  // is the authority if this ever needs to be exact. GLM-5.1 followed in April
+  // 2026. sweep: 2026-03-06.
+  'GLM-5': '2026-02-13',
+  // Apache 2.0 weights, plus official FP8 and GPTQ-Int4 quants, published at
+  // Qwen/Qwen3.5-397B-A17B on 2026-02-16 — the same day InferenceX first swept
+  // it. sweep: 2026-02-16 — day zero.
+  'Qwen-3.5-397B-A17B': '2026-02-16',
+  // Bucket covers M2.5 and M2.7, so the date is M2.5's: announced 2026-02-12
+  // with weights on Hugging Face, architecturally unchanged from M2 (230B/10B).
+  // Was 2025-10-25, which is M2's launch, not M2.5's — `model-architectures.ts`
+  // still points its sourceUrl at MiniMax-M2, which is where that came from.
+  // sweep: 2026-02-18.
+  'MiniMax-M2.5': '2026-02-12',
+  // Weights at moonshotai/Kimi-K2.5 on 2026-01-27 under a modified MIT license,
+  // 1.04T total / 32B active. sweep: 2026-02-17.
+  'Kimi-K2.5': '2026-01-27',
+  // Weights landed on Hugging Face 2025-08-05 under Apache 2.0, alongside
+  // gpt-oss-20b — OpenAI's first open-weight release since GPT-2. AWS Bedrock
+  // lists the same launch date; the model-card paper followed on 2025-08-08. Was
+  // 2025-06-13, which precedes the release by seven weeks and matches no
+  // published event. sweep: 2025-09-09.
+  'gpt-oss-120b': '2025-08-05',
+  // Announced by WeChat post and pushed to Hugging Face on 2025-05-28 under MIT
+  // — the date the model name encodes. The model card commit landed a day later,
+  // weights first. sweep: 2025-08-13.
   'DeepSeek-R1-0528': '2025-05-28',
+  // Meta published Llama 3.3 70B Instruct on 2024-12-06. sweep: 2025-08-12,
+  // shipped as workflow templates in the initial repo import.
   'Llama-3.3-70B-Instruct-FP8': '2024-12-06',
+  // Llama 3.1 was published 2024-07-23. Hidden in the model selector, but the
+  // architecture diagram still captions it. No InferenceX sweep of its own.
   'Llama-3.1-70B-Instruct-FP8-KV': '2024-07-23',
-  'gpt-oss-120b': '2025-06-13',
-  'Kimi-K2.5': '2026-01-27',
-  'Kimi-K3': '2026-06-13',
-  'MiniMax-M2.5': '2025-10-25',
-  'DeepSeek-V4-Pro': '2026-06-08',
 };
 
 /** Release date for a display model name, or null when we have no sourced date. */