@molecule/api-resource-ai-models 1.16.1 → 1.17.0

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
@@ -273,8 +273,19 @@ import type { ModelDefinition } from './types.js';
273
273
  * labeled Kimi-K3 — so kimi-k2.7-code is native-only (regions ['cn']) and
274
274
  * its US modelMap entry is gone. kimi-k3's US re-host is healthy, rates
275
275
  * unchanged.)
276
- * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
277
- * minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
276
+ * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo
277
+ * (unchanged; minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
278
+ * (re-verified 2026-10-19 across platform.minimax.io — pricing-paygo,
279
+ * models-intro, the text-generation guide, the text-chat-openai reference,
280
+ * faq/models, the model release notes and the prompt-caching page: M3 rates
281
+ * unchanged at $0.30/$1.20/cache-read $0.06; MiniMax-M3.1-Flash (released
282
+ * 2026-10-19) added at $0.15/$0.60/cache-read $0.015, 1M ctx, multimodal,
283
+ * ALWAYS thinks (thinking.type "disabled" → 400) with depth via
284
+ * reasoning_effort low|medium|high|xhigh|max (API default max); no
285
+ * cache-write line item exists for ANY MiniMax model (billing components
286
+ * are new input + cache hit + output only). M2.5/M2.7 closed to new users
287
+ * 2026-10-19 and retire 2026-11-18, with M2.x plan traffic auto-routed to
288
+ * M3.1-Flash.)
278
289
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
279
290
  * (qwen3.8-max GA'd 2026-08-03 on the pay-as-you-go international API and is
280
291
  * IN the catalog — verified 2026-08-04: flat $2/$6 per MTok on
@@ -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;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4VG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAi5EnC,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;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuWG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAi8EnC,CAAA"}
package/dist/models.js CHANGED
@@ -322,8 +322,19 @@ const CHINA_PUBLIC_HOLIDAYS = [
322
322
  * labeled Kimi-K3 — so kimi-k2.7-code is native-only (regions ['cn']) and
323
323
  * its US modelMap entry is gone. kimi-k3's US re-host is healthy, rates
324
324
  * unchanged.)
325
- * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
326
- * minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
325
+ * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo
326
+ * (unchanged; minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
327
+ * (re-verified 2026-10-19 across platform.minimax.io — pricing-paygo,
328
+ * models-intro, the text-generation guide, the text-chat-openai reference,
329
+ * faq/models, the model release notes and the prompt-caching page: M3 rates
330
+ * unchanged at $0.30/$1.20/cache-read $0.06; MiniMax-M3.1-Flash (released
331
+ * 2026-10-19) added at $0.15/$0.60/cache-read $0.015, 1M ctx, multimodal,
332
+ * ALWAYS thinks (thinking.type "disabled" → 400) with depth via
333
+ * reasoning_effort low|medium|high|xhigh|max (API default max); no
334
+ * cache-write line item exists for ANY MiniMax model (billing components
335
+ * are new input + cache hit + output only). M2.5/M2.7 closed to new users
336
+ * 2026-10-19 and retire 2026-11-18, with M2.x plan traffic auto-routed to
337
+ * M3.1-Flash.)
327
338
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
328
339
  * (qwen3.8-max GA'd 2026-08-03 on the pay-as-you-go international API and is
329
340
  * IN the catalog — verified 2026-08-04: flat $2/$6 per MTok on
@@ -2322,12 +2333,55 @@ export const MODELS = [
2322
2333
  // MiniMax
2323
2334
  // Verified: https://platform.minimax.io/docs/guides/models-intro
2324
2335
  // https://platform.minimax.io/docs/guides/pricing-paygo.md
2336
+ // https://platform.minimax.io/docs/api-reference/text-chat-openai.md
2325
2337
  // minimax-m3 (2026-06-01) is the flagship: 1M context (≥512K guaranteed),
2326
2338
  // multimodal (text+image+video in), thinking adaptive|disabled (no budget /
2327
2339
  // effort levels). Run non-thinking like the DeepSeek executors — with
2328
2340
  // thinking on, reasoning must be carried through tool loops. m2.7 thinking
2329
- // CANNOT be disabled upstream.
2341
+ // CANNOT be disabled upstream. m3.1-flash (2026-10-19) also cannot disable
2342
+ // thinking, but TUNES depth via reasoning_effort (low…max) — it is the
2343
+ // successor of the M2.x line (MiniMax routes retired M2.x plan traffic to
2344
+ // it), not of M3, which MiniMax still lists as current.
2330
2345
  // ---------------------------------------------------------------------------
2346
+ {
2347
+ id: 'minimax-m3.1-flash',
2348
+ provider: 'minimax',
2349
+ label: 'MiniMax M3.1 Flash',
2350
+ description: 'Fast multimodal MiniMax — always-on thinking, tunable effort',
2351
+ contextWindow: 1_000_000,
2352
+ // Upstream recommends 128K max_completion_tokens (hard max 512K), same as
2353
+ // M3 (text-chat-openai reference).
2354
+ maxOutputTokens: 131_072,
2355
+ // ALWAYS thinks: thinking {type:"disabled"} → 400 "requires adaptive
2356
+ // thinking", and reasoning_effort "none" is rejected too (docs + live
2357
+ // probe 2026-10-19). Depth is tuned with reasoning_effort
2358
+ // low|medium|high|xhigh|max — the bond sends adaptive + the resolved
2359
+ // level and never a disabled flag.
2360
+ supportsThinking: true,
2361
+ thinkingBudgetTokens: 16_000,
2362
+ thinkingConfigurable: true,
2363
+ supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
2364
+ // API default is max; MiniMax's own latency guidance is to LOWER the
2365
+ // effort level rather than wish thinking off — high is the balanced
2366
+ // default, xhigh/max for capability-sensitive turns.
2367
+ defaultEffortLevel: 'high',
2368
+ supportsVision: true,
2369
+ supportsPromptCaching: true,
2370
+ supportsTools: true,
2371
+ inputPricePerMTok: 0.15,
2372
+ outputPricePerMTok: 0.6,
2373
+ cacheReadPricePerMTok: 0.015,
2374
+ // No cache-write line item exists for any MiniMax model (billing is new
2375
+ // input + cache hit + output only — prompt-caching page, 2026-10-19), so
2376
+ // cache writes bill at the input rate.
2377
+ cacheWritePricePerMTok: 0.15,
2378
+ // No US re-host exists yet (open weights released 2026-10-19; not on
2379
+ // DeepInfra at verification time) — pinned to the native host like m2.5.
2380
+ regions: ['cn'],
2381
+ // Not published by MiniMax — M3's HF-template-verified cutoff carried
2382
+ // forward as the floor (this model shipped 4 months after M3).
2383
+ knowledgeCutoff: '2026-01-01',
2384
+ },
2331
2385
  {
2332
2386
  id: 'minimax-m3',
2333
2387
  provider: 'minimax',
@@ -2396,9 +2450,12 @@ export const MODELS = [
2396
2450
  us: { inputPricePerMTok: 0.25, outputPricePerMTok: 1, cacheReadPricePerMTok: 0.05 },
2397
2451
  },
2398
2452
  knowledgeCutoff: '2025-09-01',
2399
- // Superseded by minimax-m3 (same price, 1M ctx, multimodal) — kept priceable.
2453
+ // Superseded by minimax-m3.1-flash — MiniMax's own successor for the M2.x
2454
+ // line (M/Token plan traffic to M2.5/M2.7 is auto-routed there). Closed to
2455
+ // new users 2026-10-19; retired upstream 2026-11-18. Kept priceable until
2456
+ // calls stop answering.
2400
2457
  deprecatedAt: '2026-07-28',
2401
- supersededBy: 'minimax-m3',
2458
+ supersededBy: 'minimax-m3.1-flash',
2402
2459
  },
2403
2460
  {
2404
2461
  id: 'minimax-m2.5',
@@ -2421,9 +2478,11 @@ export const MODELS = [
2421
2478
  // No US re-host exists (not on DeepInfra) — pinned to native China.
2422
2479
  regions: ['cn'],
2423
2480
  knowledgeCutoff: '2025-01-01',
2424
- // Superseded by minimax-m3 (legacy upstream, still served) — kept priceable.
2481
+ // Superseded by minimax-m3.1-flash (same hand-off as m2.7 — MiniMax routes
2482
+ // retired M2.x plan traffic there; closed to new users 2026-10-19, retired
2483
+ // upstream 2026-11-18). Kept priceable until calls stop answering.
2425
2484
  deprecatedAt: '2026-03-18',
2426
- supersededBy: 'minimax-m3',
2485
+ supersededBy: 'minimax-m3.1-flash',
2427
2486
  },
2428
2487
  // ---------------------------------------------------------------------------
2429
2488
  // Alibaba (Qwen)
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@molecule/api-resource-ai-models",
3
- "version": "1.16.1",
3
+ "version": "1.17.0",
4
4
  "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",