@molecule/api-resource-ai-models 1.20.1 → 1.21.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/models.d.ts CHANGED
@@ -255,11 +255,17 @@ import type { ModelDefinition } from './types.js';
255
255
  * cataloged superseded by grok-4.7.
256
256
  * grok-4.5 still served at $2/$6, cached $0.30; grok-4.3 still served at
257
257
  * $1.25/$2.50 with the bigger 1M window; grok-code-fast-1 no longer listed)
258
- * - Chinese public holidays (DeepSeek's peak excludes them):
259
- * https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
260
- * (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
258
+ * - Chinese public holidays (DeepSeek's peak excludes them): the State
259
+ * Council's notices — 2026 at
260
+ * https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html,
261
+ * 2027 (published 2026-11-05, read 2026-11-17) at
262
+ * https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
263
+ * (see CHINA_PUBLIC_HOLIDAYS)
261
264
  * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
262
- * /updates/ (verified 2026-10-20 — the whole rate card RE-READ, unchanged:
265
+ * /updates/ (2026-11-17: the 2027 State Council holiday notice landed above —
266
+ * no DeepSeek card change; this morning's freshness-gate chunk read found the
267
+ * peak-rule sentence and every price unchanged, so the 2026-10-20
268
+ * verification stands. Earlier, verified 2026-10-20 — the whole rate card RE-READ, unchanged:
263
269
  * pro 0.66/1.98/0.022 off-peak ×2 peak, flash 0.15/0.6/0.003, peak windows
264
270
  * and the Chinese-public-holidays exclusion verbatim as before. The page is
265
271
  * now a client-rendered SPA — its HTML is an app shell and the card text
@@ -374,6 +380,24 @@ import type { ModelDefinition } from './types.js';
374
380
  * model page still describes it (docs lag the endpoint). The catalog entry
375
381
  * dropped its cn region and is now US/DeepInfra-only, whose rates re-verified
376
382
  * unchanged the same day: $0.30/$1, cache read 0.3333× input.)
383
+ * (re-verified 2026-11-17 across the provider's own pages: qwen4-max
384
+ * (released 2026-11-16) ADDED. alibabacloud.com/help/en/model-studio/
385
+ * qwen4-max + the EN model-pricing page: Singapore intl $1.5/$4.5 with
386
+ * explicit/implicit cache read $0.075 (5% of input) and every Global region
387
+ * at $1.413/$4.24 with cache read $0.071 — the ZH context-cache doc lists
388
+ * qwen4-max OUTSIDE its standard 20% cache-read table and defers to the
389
+ * console, whose per-region figures the pricing page publishes (explicit
390
+ * 5m creation 125%, 1h 200%; implicit creation 1× input). 1M ctx / 131,072
391
+ * out (262,144 thinking), Text+Image+Video input in every region, hybrid
392
+ * thinking ON by default (thinking_budget, native reasoning_effort
393
+ * low|high|xhigh exists but the bond sends the budget form), parallel tool
394
+ * calling. US = DeepInfra (Qwen/Qwen4-Max), read live the same day off
395
+ * api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
396
+ * 0.0502477× input = $0.071, no write premium, multimodal + input-video
397
+ * tags, and the host's live list carries exactly one qwen4 id (no dated
398
+ * snapshot sibling). Knowledge cutoff: not published by either host —
399
+ * best-effort estimate. qwen4-max supersedes qwen3.8-max: same
400
+ * agentic-flagship tier, cheaper at list on both hosts.)
377
401
  * - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
378
402
  * glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
379
403
  * catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
@@ -1 +1 @@
1
- {"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AA2DjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqaG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAqrFnC,CAAA"}
1
+ {"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAwEjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6bG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAmvFnC,CAAA"}
package/dist/models.js CHANGED
@@ -44,8 +44,14 @@ function dateRange(from, to) {
44
44
  * 2026-09-23). Those windows are 09:00–12:00 and 14:00–18:00 Beijing time, so
45
45
  * the Beijing holiday date IS the UTC date of every window.
46
46
  *
47
- * Source — the State Council's notice for 2026 (国办发明电〔2025〕7号), read
48
- * 2026-09-23 at https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html.
47
+ * Sources — the State Council's notices, 放假 ranges verbatim:
48
+ * - 2026 (国办发明电〔2025〕7号), read 2026-09-23 at
49
+ * https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
50
+ * - 2027 (published 2026-11-05), read 2026-11-17 at
51
+ * https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
52
+ * (People's Daily reproduction of the 新华社 dispatch; cross-checked against
53
+ * news.ifeng.com/c/8xoAMrtRK7T — all seven clauses and the 2026-11-05 date
54
+ * identical, and every 放假 weekday annotation matches the 2027 calendar).
49
55
  * Make-up working days (调休 weekends) need no entry: the peak is Mon–Fri only.
50
56
  *
51
57
  * ADD NEXT YEAR'S DATES when the State Council publishes them (each November);
@@ -59,6 +65,13 @@ const CHINA_PUBLIC_HOLIDAYS = [
59
65
  ...dateRange('2026-06-19', '2026-06-21'), // 端午节
60
66
  ...dateRange('2026-09-25', '2026-09-27'), // 中秋节
61
67
  ...dateRange('2026-10-01', '2026-10-07'), // 国庆节
68
+ ...dateRange('2027-01-01', '2027-01-03'), // 元旦
69
+ ...dateRange('2027-02-05', '2027-02-13'), // 春节 (除夕 to 正月初八, 9 days)
70
+ ...dateRange('2027-04-03', '2027-04-05'), // 清明节
71
+ ...dateRange('2027-05-01', '2027-05-05'), // 劳动节
72
+ ...dateRange('2027-06-09', '2027-06-09'), // 端午节 (single day)
73
+ ...dateRange('2027-09-15', '2027-09-15'), // 中秋节 (single day)
74
+ ...dateRange('2027-10-01', '2027-10-07'), // 国庆节
62
75
  ];
63
76
  /**
64
77
  * All available AI models, grouped by provider, ordered from most to least capable.
@@ -307,11 +320,17 @@ const CHINA_PUBLIC_HOLIDAYS = [
307
320
  * cataloged superseded by grok-4.7.
308
321
  * grok-4.5 still served at $2/$6, cached $0.30; grok-4.3 still served at
309
322
  * $1.25/$2.50 with the bigger 1M window; grok-code-fast-1 no longer listed)
310
- * - Chinese public holidays (DeepSeek's peak excludes them):
311
- * https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
312
- * (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
323
+ * - Chinese public holidays (DeepSeek's peak excludes them): the State
324
+ * Council's notices — 2026 at
325
+ * https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html,
326
+ * 2027 (published 2026-11-05, read 2026-11-17) at
327
+ * https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
328
+ * (see CHINA_PUBLIC_HOLIDAYS)
313
329
  * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
314
- * /updates/ (verified 2026-10-20 — the whole rate card RE-READ, unchanged:
330
+ * /updates/ (2026-11-17: the 2027 State Council holiday notice landed above —
331
+ * no DeepSeek card change; this morning's freshness-gate chunk read found the
332
+ * peak-rule sentence and every price unchanged, so the 2026-10-20
333
+ * verification stands. Earlier, verified 2026-10-20 — the whole rate card RE-READ, unchanged:
315
334
  * pro 0.66/1.98/0.022 off-peak ×2 peak, flash 0.15/0.6/0.003, peak windows
316
335
  * and the Chinese-public-holidays exclusion verbatim as before. The page is
317
336
  * now a client-rendered SPA — its HTML is an app shell and the card text
@@ -426,6 +445,24 @@ const CHINA_PUBLIC_HOLIDAYS = [
426
445
  * model page still describes it (docs lag the endpoint). The catalog entry
427
446
  * dropped its cn region and is now US/DeepInfra-only, whose rates re-verified
428
447
  * unchanged the same day: $0.30/$1, cache read 0.3333× input.)
448
+ * (re-verified 2026-11-17 across the provider's own pages: qwen4-max
449
+ * (released 2026-11-16) ADDED. alibabacloud.com/help/en/model-studio/
450
+ * qwen4-max + the EN model-pricing page: Singapore intl $1.5/$4.5 with
451
+ * explicit/implicit cache read $0.075 (5% of input) and every Global region
452
+ * at $1.413/$4.24 with cache read $0.071 — the ZH context-cache doc lists
453
+ * qwen4-max OUTSIDE its standard 20% cache-read table and defers to the
454
+ * console, whose per-region figures the pricing page publishes (explicit
455
+ * 5m creation 125%, 1h 200%; implicit creation 1× input). 1M ctx / 131,072
456
+ * out (262,144 thinking), Text+Image+Video input in every region, hybrid
457
+ * thinking ON by default (thinking_budget, native reasoning_effort
458
+ * low|high|xhigh exists but the bond sends the budget form), parallel tool
459
+ * calling. US = DeepInfra (Qwen/Qwen4-Max), read live the same day off
460
+ * api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
461
+ * 0.0502477× input = $0.071, no write premium, multimodal + input-video
462
+ * tags, and the host's live list carries exactly one qwen4 id (no dated
463
+ * snapshot sibling). Knowledge cutoff: not published by either host —
464
+ * best-effort estimate. qwen4-max supersedes qwen3.8-max: same
465
+ * agentic-flagship tier, cheaper at list on both hosts.)
429
466
  * - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
430
467
  * glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
431
468
  * catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
@@ -2796,20 +2833,78 @@ export const MODELS = [
2796
2833
  // Verified: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
2797
2834
  // https://www.alibabacloud.com/help/en/model-studio/qwen-coder
2798
2835
  // https://openrouter.ai/qwen/qwen3.7-max
2799
- // qwen3.8-max (2026-08-03) is the agentic flagship, succeeding qwen3.7-max —
2800
- // Alibaba's own Qwen-Coder docs now recommend the general-purpose models over
2801
- // Qwen-Coder. Their thinking
2802
- // uses enable_thinking (default ON for the 3.7 series) + thinking_budget
2803
- // (token cap) — a real budget param, so effort scales the budget. The
2804
- // qwen3-coder models are NON-thinking (previous catalog entry was wrong).
2836
+ // qwen4-max (2026-11-16) is the agentic flagship, succeeding qwen3.8-max —
2837
+ // multimodal (Text+Image+Video input) on BOTH hosts this generation. Hybrid
2838
+ // thinking uses enable_thinking (default ON) + thinking_budget (token cap) —
2839
+ // a real budget param, so effort scales the budget. (Qwen4-Max also exposes a
2840
+ // native reasoning_effort low|high|xhigh, but the alibaba bond speaks only
2841
+ // the budget form, so the catalog models effort as scaled budgets like the
2842
+ // 3.x flagships.) The qwen3-coder models are NON-thinking.
2805
2843
  // Prices are DashScope international list rates (the bond calls DashScope,
2806
- // not OpenRouter; a 50%-off promo currently applies — billed at list).
2844
+ // not OpenRouter; bill at list).
2807
2845
  // ---------------------------------------------------------------------------
2808
- // qwen3.8-max (GA 2026-08-03) succeeds qwen3.7-max as the agentic flagship,
2846
+ // qwen4-max (2026-11-16) succeeds qwen3.8-max as the agentic flagship,
2847
+ // priced BELOW it at $1.5/$4.5 on the Singapore international card (every
2848
+ // other native region and the US re-host bill $1.413/$4.24). Same hybrid
2849
+ // thinking mechanism as the 3.x series. Cache rates are published per region
2850
+ // on the model-pricing page — see the Sources block for the citation.
2851
+ {
2852
+ id: 'qwen4-max',
2853
+ provider: 'alibaba',
2854
+ label: 'Qwen4 Max',
2855
+ description: 'Alibaba agentic flagship — 1M context, multimodal, hybrid thinking',
2856
+ // Model page (verified 2026-11-17): context 1,000,000; max input 991,808
2857
+ // (983,616 thinking); max output 131,072 (262,144 in thinking mode); max
2858
+ // chain-of-thought 262,144. Kept at the non-thinking output figure —
2859
+ // understating only shortens completions.
2860
+ contextWindow: 1_000_000,
2861
+ maxOutputTokens: 131_072,
2862
+ supportsThinking: true,
2863
+ thinkingBudgetTokens: 8_000,
2864
+ thinkingConfigurable: true,
2865
+ supportedEffortLevels: ['4K', '8K', '16K', '32K'],
2866
+ defaultEffortLevel: '8K',
2867
+ effortBudgetTokens: { '4K': 4000, '8K': 8000, '16K': 16000, '32K': 32000 },
2868
+ // Multimodal on BOTH hosts this generation: the model page lists
2869
+ // Text/Image/Video input in every region, and the DeepInfra re-host's own
2870
+ // record tags the id multimodal + input-video (Qwen3.8-Max's re-host
2871
+ // carried no such tags — that is why vision stayed false there).
2872
+ supportsVision: true,
2873
+ supportsPromptCaching: true,
2874
+ supportsTools: true,
2875
+ // Singapore international list card (model page + model-pricing page,
2876
+ // verified 2026-11-17). The ZH context-cache doc lists qwen4-max among the
2877
+ // models OUTSIDE the standard 20% cache-read table and defers to the
2878
+ // console, whose figures the pricing page publishes per region: cache read
2879
+ // $0.075 in Singapore (5% of input; $0.071 in every Global region).
2880
+ // Implicit creation is billed at plain input (write 1×); the explicit 5m
2881
+ // creation premium is 125% = $1.875, which implicit-mode billing never
2882
+ // charges.
2883
+ inputPricePerMTok: 1.5,
2884
+ outputPricePerMTok: 4.5,
2885
+ cacheReadPricePerMTok: 0.075,
2886
+ cacheWritePricePerMTok: 1.5,
2887
+ regions: ['us', 'cn'],
2888
+ // US = DeepInfra (Qwen/Qwen4-Max), verified 2026-11-17 against
2889
+ // api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
2890
+ // 0.0502477× input = $0.071, no write premium (omitted → region input).
2891
+ // The host's live model list carries exactly one qwen4 id — no dated
2892
+ // snapshot sibling — and the entry is not deprecated.
2893
+ regionPricing: {
2894
+ us: { inputPricePerMTok: 1.413, outputPricePerMTok: 4.24, cacheReadPricePerMTok: 0.071 },
2895
+ },
2896
+ // Not published by Alibaba — best-effort estimate for the Qwen4 generation
2897
+ // (released 2026-11-16).
2898
+ knowledgeCutoff: '2026-08-01',
2899
+ },
2900
+ // qwen3.8-max (GA 2026-08-03) succeeded qwen3.7-max as the agentic flagship,
2809
2901
  // priced BELOW it at $2/$6 (intl CNY 14.988/44.965, same fixed conversion).
2810
2902
  // Same hybrid thinking mechanism as the 3.7 series (enable_thinking default
2811
2903
  // ON + thinking_budget; preserve_thinking supported). Context cache uses the
2812
2904
  // standard implicit rates — see the Sources block for the ZH-doc citation.
2905
+ // Superseded by qwen4-max (2026-11-16): same tier at a LOWER list price on
2906
+ // both hosts. Still served upstream, so it stays priceable — just not
2907
+ // offered.
2813
2908
  {
2814
2909
  id: 'qwen3.8-max',
2815
2910
  provider: 'alibaba',
@@ -2856,6 +2951,8 @@ export const MODELS = [
2856
2951
  },
2857
2952
  // Not published by Alibaba — best-effort estimate.
2858
2953
  knowledgeCutoff: '2026-04-01',
2954
+ deprecatedAt: '2026-11-16',
2955
+ supersededBy: 'qwen4-max',
2859
2956
  },
2860
2957
  {
2861
2958
  id: 'qwen3.8-flash',
@@ -2933,11 +3030,13 @@ export const MODELS = [
2933
3030
  },
2934
3031
  // Not published by Alibaba — best-effort estimate.
2935
3032
  knowledgeCutoff: '2026-01-01',
2936
- // Superseded by qwen3.8-max (GA 2026-08-03): same tier and mechanism, and
2937
- // CHEAPER at list ($2/$6 vs $2.50/$7.50). Still served upstream (the 50%-off
2938
- // promo runs on this id), so it stays priceable — it is just not offered.
3033
+ // Superseded (GA 2026-08-03) and now pointed one hop forward at qwen4-max
3034
+ // (2026-11-16): same tier and mechanism, each generation CHEAPER at list
3035
+ // ($2/$6 then $1.5/$4.5, vs this id's $2.50/$7.50). Still served upstream
3036
+ // (the 50%-off promo runs on this id), so it stays priceable — it is just
3037
+ // not offered.
2939
3038
  deprecatedAt: '2026-08-03',
2940
- supersededBy: 'qwen3.8-max',
3039
+ supersededBy: 'qwen4-max',
2941
3040
  },
2942
3041
  {
2943
3042
  id: 'qwen3-coder-plus',
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "@molecule/api-resource-ai-models",
3
- "version": "1.20.1",
4
- "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
3
+ "version": "1.21.0",
4
+ "description": "Server-side AI model catalog plus GET /ai/models endpoint",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",
7
7
  "types": "dist/index.d.ts",
@@ -28,8 +28,8 @@
28
28
  "license": "Apache-2.0",
29
29
  "devDependencies": {
30
30
  "@molecule/api-bond": "1.0.4",
31
- "@molecule/api-i18n": "1.0.5",
32
- "@molecule/api-resource": "1.1.2",
31
+ "@molecule/api-i18n": "1.0.6",
32
+ "@molecule/api-resource": "1.1.3",
33
33
  "@types/node": "26.6.5",
34
34
  "typescript": "6.0.3",
35
35
  "vitest": "4.1.12"