pi-free 1.0.8 → 2.0.0

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.
Files changed (63) hide show
  1. package/CHANGELOG.md +107 -1
  2. package/README.md +95 -46
  3. package/config.ts +165 -120
  4. package/constants.ts +22 -61
  5. package/index.ts +186 -0
  6. package/lib/json-persistence.ts +11 -10
  7. package/lib/logger.ts +2 -2
  8. package/lib/model-enhancer.ts +20 -20
  9. package/lib/open-browser.ts +41 -0
  10. package/lib/provider-cache.ts +106 -0
  11. package/lib/registry.ts +144 -0
  12. package/package.json +67 -82
  13. package/provider-factory.ts +25 -41
  14. package/provider-failover/benchmark-lookup.ts +247 -0
  15. package/provider-failover/benchmarks-chunk-0.ts +2010 -0
  16. package/provider-failover/benchmarks-chunk-1.ts +1988 -0
  17. package/provider-failover/benchmarks-chunk-2.ts +2010 -0
  18. package/provider-failover/benchmarks-chunk-3.ts +2010 -0
  19. package/provider-failover/benchmarks-chunk-4.ts +1969 -0
  20. package/provider-failover/hardcoded-benchmarks.ts +22 -10025
  21. package/provider-helper.ts +38 -37
  22. package/providers/{cline-auth.ts → cline/cline-auth.ts} +2 -2
  23. package/providers/cline/cline-models.ts +128 -0
  24. package/providers/{cline.ts → cline/cline.ts} +300 -257
  25. package/providers/cloudflare/cloudflare.ts +368 -0
  26. package/providers/dynamic-built-in/index.ts +513 -0
  27. package/providers/{kilo-auth.ts → kilo/kilo-auth.ts} +3 -20
  28. package/providers/{kilo-models.ts → kilo/kilo-models.ts} +2 -2
  29. package/providers/kilo/kilo.ts +235 -0
  30. package/providers/{modal.ts → modal/modal.ts} +4 -3
  31. package/providers/{nvidia.ts → nvidia/nvidia.ts} +152 -113
  32. package/providers/ollama/ollama.ts +172 -0
  33. package/providers/opencode-session.ts +34 -34
  34. package/providers/{qwen-auth.ts → qwen/qwen-auth.ts} +24 -40
  35. package/providers/{qwen-models.ts → qwen/qwen-models.ts} +101 -95
  36. package/providers/qwen/qwen.ts +202 -0
  37. package/provider-failover/auto-switch.ts +0 -350
  38. package/provider-failover/errors.ts +0 -275
  39. package/provider-failover/index.ts +0 -238
  40. package/providers/cline-models.ts +0 -77
  41. package/providers/factory.ts +0 -125
  42. package/providers/fireworks.ts +0 -49
  43. package/providers/go.ts +0 -216
  44. package/providers/kilo.ts +0 -146
  45. package/providers/mistral.ts +0 -144
  46. package/providers/ollama.ts +0 -113
  47. package/providers/openrouter.ts +0 -175
  48. package/providers/qwen.ts +0 -127
  49. package/providers/zen.ts +0 -371
  50. package/usage/commands.ts +0 -17
  51. package/usage/cumulative.ts +0 -193
  52. package/usage/formatters.ts +0 -115
  53. package/usage/index.ts +0 -46
  54. package/usage/limits.ts +0 -148
  55. package/usage/metrics.ts +0 -222
  56. package/usage/sessions.ts +0 -355
  57. package/usage/store.ts +0 -99
  58. package/usage/tracking.ts +0 -329
  59. package/usage/types.ts +0 -26
  60. package/usage/widget.ts +0 -90
  61. package/widget/data.ts +0 -113
  62. package/widget/format.ts +0 -26
  63. package/widget/render.ts +0 -117
@@ -0,0 +1,1969 @@
1
+ // Auto-generated benchmark data chunk 4
2
+ // Models: lfm-40b .. seed-oss-36b-instruct (89 entries)
3
+ // DO NOT EDIT MANUALLY — generated by scripts/update-benchmarks.ts
4
+
5
+ import type { HardcodedBenchmark } from "./hardcoded-benchmarks.ts";
6
+
7
+ export const BENCHMARKS_CHUNK_4: Record<string, HardcodedBenchmark> = {
8
+ "lfm-40b": {
9
+ // AA Intelligence Index (composite score)
10
+ intelligenceIndex: 8.8,
11
+ normalizedScore: 13,
12
+
13
+ // AA specific benchmarks
14
+ codingIndex: undefined,
15
+ mathIndex: undefined,
16
+
17
+ // Academic benchmarks
18
+ mmluPro: 0.425,
19
+ gpqa: 0.327,
20
+ hle: 0.049,
21
+
22
+ // Capabilities
23
+ contextWindow: 8192,
24
+ supportsReasoning: false,
25
+ supportsVision: false,
26
+
27
+ // Metadata
28
+ lastUpdated: "2026-04-06",
29
+ },
30
+ "lfm2-1.2b": {
31
+ // AA Intelligence Index (composite score)
32
+ intelligenceIndex: 6.3,
33
+ normalizedScore: 9,
34
+
35
+ // AA specific benchmarks
36
+ codingIndex: 0.8,
37
+ mathIndex: 3.3,
38
+
39
+ // Academic benchmarks
40
+ mmluPro: 0.257,
41
+ gpqa: 0.228,
42
+ hle: 0.057,
43
+
44
+ // Capabilities
45
+ contextWindow: 8192,
46
+ supportsReasoning: false,
47
+ supportsVision: false,
48
+
49
+ // Metadata
50
+ lastUpdated: "2026-04-06",
51
+ },
52
+ "solar-mini": {
53
+ // AA Intelligence Index (composite score)
54
+ intelligenceIndex: 11.9,
55
+ normalizedScore: 17,
56
+
57
+ // AA specific benchmarks
58
+ codingIndex: undefined,
59
+ mathIndex: undefined,
60
+
61
+ // Academic benchmarks
62
+ mmluPro: undefined,
63
+ gpqa: undefined,
64
+ hle: undefined,
65
+
66
+ // Capabilities
67
+ contextWindow: 8192,
68
+ supportsReasoning: false,
69
+ supportsVision: false,
70
+
71
+ // Metadata
72
+ lastUpdated: "2026-04-06",
73
+ },
74
+ "solar-pro-2-preview-non-reasoning": {
75
+ // AA Intelligence Index (composite score)
76
+ intelligenceIndex: 16.0,
77
+ normalizedScore: 23,
78
+
79
+ // AA specific benchmarks
80
+ codingIndex: undefined,
81
+ mathIndex: undefined,
82
+
83
+ // Academic benchmarks
84
+ mmluPro: 0.725,
85
+ gpqa: 0.544,
86
+ hle: 0.038,
87
+
88
+ // Capabilities
89
+ contextWindow: 8192,
90
+ supportsReasoning: false,
91
+ supportsVision: false,
92
+
93
+ // Metadata
94
+ lastUpdated: "2026-04-06",
95
+ },
96
+ "solar-pro-2-preview-reasoning": {
97
+ // AA Intelligence Index (composite score)
98
+ intelligenceIndex: 18.8,
99
+ normalizedScore: 27,
100
+
101
+ // AA specific benchmarks
102
+ codingIndex: undefined,
103
+ mathIndex: undefined,
104
+
105
+ // Academic benchmarks
106
+ mmluPro: 0.768,
107
+ gpqa: 0.578,
108
+ hle: 0.057,
109
+
110
+ // Capabilities
111
+ contextWindow: 8192,
112
+ supportsReasoning: false,
113
+ supportsVision: false,
114
+
115
+ // Metadata
116
+ lastUpdated: "2026-04-06",
117
+ },
118
+ "dbrx-instruct": {
119
+ // AA Intelligence Index (composite score)
120
+ intelligenceIndex: 8.3,
121
+ normalizedScore: 12,
122
+
123
+ // AA specific benchmarks
124
+ codingIndex: undefined,
125
+ mathIndex: undefined,
126
+
127
+ // Academic benchmarks
128
+ mmluPro: 0.397,
129
+ gpqa: 0.331,
130
+ hle: 0.066,
131
+
132
+ // Capabilities
133
+ contextWindow: 8192,
134
+ supportsReasoning: false,
135
+ supportsVision: false,
136
+
137
+ // Metadata
138
+ lastUpdated: "2026-04-06",
139
+ },
140
+ "minimax-m2.5": {
141
+ // AA Intelligence Index (composite score)
142
+ intelligenceIndex: 41.9,
143
+ normalizedScore: 60,
144
+
145
+ // AA specific benchmarks
146
+ codingIndex: 37.4,
147
+ mathIndex: undefined,
148
+
149
+ // Academic benchmarks
150
+ mmluPro: undefined,
151
+ gpqa: 0.848,
152
+ hle: 0.191,
153
+
154
+ // Capabilities
155
+ contextWindow: 8192,
156
+ supportsReasoning: false,
157
+ supportsVision: false,
158
+
159
+ // Metadata
160
+ lastUpdated: "2026-04-06",
161
+ },
162
+ "minimax-m2.1": {
163
+ // AA Intelligence Index (composite score)
164
+ intelligenceIndex: 39.4,
165
+ normalizedScore: 56,
166
+
167
+ // AA specific benchmarks
168
+ codingIndex: 32.8,
169
+ mathIndex: 82.7,
170
+
171
+ // Academic benchmarks
172
+ mmluPro: 0.875,
173
+ gpqa: 0.83,
174
+ hle: 0.222,
175
+
176
+ // Capabilities
177
+ contextWindow: 8192,
178
+ supportsReasoning: false,
179
+ supportsVision: false,
180
+
181
+ // Metadata
182
+ lastUpdated: "2026-04-06",
183
+ },
184
+ "minimax-m1-80k": {
185
+ // AA Intelligence Index (composite score)
186
+ intelligenceIndex: 24.4,
187
+ normalizedScore: 35,
188
+
189
+ // AA specific benchmarks
190
+ codingIndex: 14.5,
191
+ mathIndex: 61.0,
192
+
193
+ // Academic benchmarks
194
+ mmluPro: 0.816,
195
+ gpqa: 0.697,
196
+ hle: 0.082,
197
+
198
+ // Capabilities
199
+ contextWindow: 8192,
200
+ supportsReasoning: false,
201
+ supportsVision: false,
202
+
203
+ // Metadata
204
+ lastUpdated: "2026-04-06",
205
+ },
206
+ "minimax-m2": {
207
+ // AA Intelligence Index (composite score)
208
+ intelligenceIndex: 36.1,
209
+ normalizedScore: 52,
210
+
211
+ // AA specific benchmarks
212
+ codingIndex: 29.2,
213
+ mathIndex: 78.3,
214
+
215
+ // Academic benchmarks
216
+ mmluPro: 0.82,
217
+ gpqa: 0.777,
218
+ hle: 0.125,
219
+
220
+ // Capabilities
221
+ contextWindow: 8192,
222
+ supportsReasoning: false,
223
+ supportsVision: false,
224
+
225
+ // Metadata
226
+ lastUpdated: "2026-04-06",
227
+ },
228
+ "minimax-m1-40k": {
229
+ // AA Intelligence Index (composite score)
230
+ intelligenceIndex: 20.9,
231
+ normalizedScore: 30,
232
+
233
+ // AA specific benchmarks
234
+ codingIndex: 14.1,
235
+ mathIndex: 13.7,
236
+
237
+ // Academic benchmarks
238
+ mmluPro: 0.808,
239
+ gpqa: 0.682,
240
+ hle: 0.075,
241
+
242
+ // Capabilities
243
+ contextWindow: 8192,
244
+ supportsReasoning: false,
245
+ supportsVision: false,
246
+
247
+ // Metadata
248
+ lastUpdated: "2026-04-06",
249
+ },
250
+ "kimi-k2-thinking": {
251
+ // AA Intelligence Index (composite score)
252
+ intelligenceIndex: 40.9,
253
+ normalizedScore: 58,
254
+
255
+ // AA specific benchmarks
256
+ codingIndex: 34.8,
257
+ mathIndex: 94.7,
258
+
259
+ // Academic benchmarks
260
+ mmluPro: 0.848,
261
+ gpqa: 0.838,
262
+ hle: 0.223,
263
+
264
+ // Capabilities
265
+ contextWindow: 8192,
266
+ supportsReasoning: false,
267
+ supportsVision: false,
268
+
269
+ // Metadata
270
+ lastUpdated: "2026-04-06",
271
+ },
272
+ "kimi-k2-0905": {
273
+ // AA Intelligence Index (composite score)
274
+ intelligenceIndex: 30.9,
275
+ normalizedScore: 44,
276
+
277
+ // AA specific benchmarks
278
+ codingIndex: 25.9,
279
+ mathIndex: 57.3,
280
+
281
+ // Academic benchmarks
282
+ mmluPro: 0.819,
283
+ gpqa: 0.767,
284
+ hle: 0.063,
285
+
286
+ // Capabilities
287
+ contextWindow: 8192,
288
+ supportsReasoning: false,
289
+ supportsVision: false,
290
+
291
+ // Metadata
292
+ lastUpdated: "2026-04-06",
293
+ },
294
+ "kimi-k2": {
295
+ // AA Intelligence Index (composite score)
296
+ intelligenceIndex: 26.3,
297
+ normalizedScore: 38,
298
+
299
+ // AA specific benchmarks
300
+ codingIndex: 22.1,
301
+ mathIndex: 57.0,
302
+
303
+ // Academic benchmarks
304
+ mmluPro: 0.824,
305
+ gpqa: 0.766,
306
+ hle: 0.07,
307
+
308
+ // Capabilities
309
+ contextWindow: 8192,
310
+ supportsReasoning: false,
311
+ supportsVision: false,
312
+
313
+ // Metadata
314
+ lastUpdated: "2026-04-06",
315
+ },
316
+ "llama-3.1-tulu3-405b": {
317
+ // AA Intelligence Index (composite score)
318
+ intelligenceIndex: 14.1,
319
+ normalizedScore: 20,
320
+
321
+ // AA specific benchmarks
322
+ codingIndex: undefined,
323
+ mathIndex: undefined,
324
+
325
+ // Academic benchmarks
326
+ mmluPro: 0.716,
327
+ gpqa: 0.516,
328
+ hle: 0.035,
329
+
330
+ // Capabilities
331
+ contextWindow: 8192,
332
+ supportsReasoning: false,
333
+ supportsVision: false,
334
+
335
+ // Metadata
336
+ lastUpdated: "2026-04-06",
337
+ },
338
+ "olmo-2-7b": {
339
+ // AA Intelligence Index (composite score)
340
+ intelligenceIndex: 9.3,
341
+ normalizedScore: 13,
342
+
343
+ // AA specific benchmarks
344
+ codingIndex: 1.2,
345
+ mathIndex: 0.7,
346
+
347
+ // Academic benchmarks
348
+ mmluPro: 0.282,
349
+ gpqa: 0.288,
350
+ hle: 0.055,
351
+
352
+ // Capabilities
353
+ contextWindow: 8192,
354
+ supportsReasoning: false,
355
+ supportsVision: false,
356
+
357
+ // Metadata
358
+ lastUpdated: "2026-04-06",
359
+ },
360
+ "olmo-2-32b": {
361
+ // AA Intelligence Index (composite score)
362
+ intelligenceIndex: 10.6,
363
+ normalizedScore: 15,
364
+
365
+ // AA specific benchmarks
366
+ codingIndex: 2.7,
367
+ mathIndex: 3.3,
368
+
369
+ // Academic benchmarks
370
+ mmluPro: 0.511,
371
+ gpqa: 0.328,
372
+ hle: 0.037,
373
+
374
+ // Capabilities
375
+ contextWindow: 8192,
376
+ supportsReasoning: false,
377
+ supportsVision: false,
378
+
379
+ // Metadata
380
+ lastUpdated: "2026-04-06",
381
+ },
382
+ "olmo-3-32b-think": {
383
+ // AA Intelligence Index (composite score)
384
+ intelligenceIndex: 12.1,
385
+ normalizedScore: 17,
386
+
387
+ // AA specific benchmarks
388
+ codingIndex: 10.5,
389
+ mathIndex: 73.7,
390
+
391
+ // Academic benchmarks
392
+ mmluPro: 0.759,
393
+ gpqa: 0.61,
394
+ hle: 0.059,
395
+
396
+ // Capabilities
397
+ contextWindow: 8192,
398
+ supportsReasoning: false,
399
+ supportsVision: false,
400
+
401
+ // Metadata
402
+ lastUpdated: "2026-04-06",
403
+ },
404
+ "granite-3.3-8b-non-reasoning": {
405
+ // AA Intelligence Index (composite score)
406
+ intelligenceIndex: 7.0,
407
+ normalizedScore: 10,
408
+
409
+ // AA specific benchmarks
410
+ codingIndex: 3.4,
411
+ mathIndex: 6.7,
412
+
413
+ // Academic benchmarks
414
+ mmluPro: 0.468,
415
+ gpqa: 0.338,
416
+ hle: 0.042,
417
+
418
+ // Capabilities
419
+ contextWindow: 8192,
420
+ supportsReasoning: false,
421
+ supportsVision: false,
422
+
423
+ // Metadata
424
+ lastUpdated: "2026-04-06",
425
+ },
426
+ "reka-flash-sep-24": {
427
+ // AA Intelligence Index (composite score)
428
+ intelligenceIndex: 12.0,
429
+ normalizedScore: 17,
430
+
431
+ // AA specific benchmarks
432
+ codingIndex: undefined,
433
+ mathIndex: undefined,
434
+
435
+ // Academic benchmarks
436
+ mmluPro: undefined,
437
+ gpqa: undefined,
438
+ hle: undefined,
439
+
440
+ // Capabilities
441
+ contextWindow: 8192,
442
+ supportsReasoning: false,
443
+ supportsVision: false,
444
+
445
+ // Metadata
446
+ lastUpdated: "2026-04-06",
447
+ },
448
+ "hermes-3---llama-3.1-70b": {
449
+ // AA Intelligence Index (composite score)
450
+ intelligenceIndex: 10.6,
451
+ normalizedScore: 15,
452
+
453
+ // AA specific benchmarks
454
+ codingIndex: undefined,
455
+ mathIndex: undefined,
456
+
457
+ // Academic benchmarks
458
+ mmluPro: 0.571,
459
+ gpqa: 0.401,
460
+ hle: 0.041,
461
+
462
+ // Capabilities
463
+ contextWindow: 8192,
464
+ supportsReasoning: false,
465
+ supportsVision: false,
466
+
467
+ // Metadata
468
+ lastUpdated: "2026-04-06",
469
+ },
470
+ "mimo-v2-flash-reasoning": {
471
+ // AA Intelligence Index (composite score)
472
+ intelligenceIndex: 39.2,
473
+ normalizedScore: 56,
474
+
475
+ // AA specific benchmarks
476
+ codingIndex: 31.8,
477
+ mathIndex: 96.3,
478
+
479
+ // Academic benchmarks
480
+ mmluPro: 0.843,
481
+ gpqa: 0.846,
482
+ hle: 0.211,
483
+
484
+ // Capabilities
485
+ contextWindow: 8192,
486
+ supportsReasoning: false,
487
+ supportsVision: false,
488
+
489
+ // Metadata
490
+ lastUpdated: "2026-04-06",
491
+ },
492
+ "sarvam-m-reasoning": {
493
+ // AA Intelligence Index (composite score)
494
+ intelligenceIndex: 8.4,
495
+ normalizedScore: 12,
496
+
497
+ // AA specific benchmarks
498
+ codingIndex: 7.5,
499
+ mathIndex: undefined,
500
+
501
+ // Academic benchmarks
502
+ mmluPro: 0.696,
503
+ gpqa: 0.416,
504
+ hle: 0.033,
505
+
506
+ // Capabilities
507
+ contextWindow: 8192,
508
+ supportsReasoning: false,
509
+ supportsVision: false,
510
+
511
+ // Metadata
512
+ lastUpdated: "2026-04-06",
513
+ },
514
+ "glm-4.6-non-reasoning": {
515
+ // AA Intelligence Index (composite score)
516
+ intelligenceIndex: 30.2,
517
+ normalizedScore: 43,
518
+
519
+ // AA specific benchmarks
520
+ codingIndex: 30.2,
521
+ mathIndex: 44.3,
522
+
523
+ // Academic benchmarks
524
+ mmluPro: 0.784,
525
+ gpqa: 0.632,
526
+ hle: 0.052,
527
+
528
+ // Capabilities
529
+ contextWindow: 8192,
530
+ supportsReasoning: false,
531
+ supportsVision: false,
532
+
533
+ // Metadata
534
+ lastUpdated: "2026-04-06",
535
+ },
536
+ "glm-4.7-reasoning": {
537
+ // AA Intelligence Index (composite score)
538
+ intelligenceIndex: 42.1,
539
+ normalizedScore: 60,
540
+
541
+ // AA specific benchmarks
542
+ codingIndex: 36.3,
543
+ mathIndex: 95.0,
544
+
545
+ // Academic benchmarks
546
+ mmluPro: 0.856,
547
+ gpqa: 0.859,
548
+ hle: 0.251,
549
+
550
+ // Capabilities
551
+ contextWindow: 8192,
552
+ supportsReasoning: false,
553
+ supportsVision: false,
554
+
555
+ // Metadata
556
+ lastUpdated: "2026-04-06",
557
+ },
558
+ "glm-4.7-flash-non-reasoning": {
559
+ // AA Intelligence Index (composite score)
560
+ intelligenceIndex: 22.1,
561
+ normalizedScore: 32,
562
+
563
+ // AA specific benchmarks
564
+ codingIndex: 11.0,
565
+ mathIndex: undefined,
566
+
567
+ // Academic benchmarks
568
+ mmluPro: undefined,
569
+ gpqa: 0.452,
570
+ hle: 0.049,
571
+
572
+ // Capabilities
573
+ contextWindow: 8192,
574
+ supportsReasoning: false,
575
+ supportsVision: false,
576
+
577
+ // Metadata
578
+ lastUpdated: "2026-04-06",
579
+ },
580
+ "glm-4.7-non-reasoning": {
581
+ // AA Intelligence Index (composite score)
582
+ intelligenceIndex: 34.2,
583
+ normalizedScore: 49,
584
+
585
+ // AA specific benchmarks
586
+ codingIndex: 32.0,
587
+ mathIndex: 48.0,
588
+
589
+ // Academic benchmarks
590
+ mmluPro: 0.794,
591
+ gpqa: 0.664,
592
+ hle: 0.061,
593
+
594
+ // Capabilities
595
+ contextWindow: 8192,
596
+ supportsReasoning: false,
597
+ supportsVision: false,
598
+
599
+ // Metadata
600
+ lastUpdated: "2026-04-06",
601
+ },
602
+ "glm-4.5v-non-reasoning": {
603
+ // AA Intelligence Index (composite score)
604
+ intelligenceIndex: 12.7,
605
+ normalizedScore: 18,
606
+
607
+ // AA specific benchmarks
608
+ codingIndex: 10.8,
609
+ mathIndex: 15.3,
610
+
611
+ // Academic benchmarks
612
+ mmluPro: 0.751,
613
+ gpqa: 0.573,
614
+ hle: 0.036,
615
+
616
+ // Capabilities
617
+ contextWindow: 8192,
618
+ supportsReasoning: false,
619
+ supportsVision: false,
620
+
621
+ // Metadata
622
+ lastUpdated: "2026-04-06",
623
+ },
624
+ "glm-4.5-reasoning": {
625
+ // AA Intelligence Index (composite score)
626
+ intelligenceIndex: 26.4,
627
+ normalizedScore: 38,
628
+
629
+ // AA specific benchmarks
630
+ codingIndex: 26.3,
631
+ mathIndex: 73.7,
632
+
633
+ // Academic benchmarks
634
+ mmluPro: 0.835,
635
+ gpqa: 0.782,
636
+ hle: 0.122,
637
+
638
+ // Capabilities
639
+ contextWindow: 8192,
640
+ supportsReasoning: false,
641
+ supportsVision: false,
642
+
643
+ // Metadata
644
+ lastUpdated: "2026-04-06",
645
+ },
646
+ "glm-4.6-reasoning": {
647
+ // AA Intelligence Index (composite score)
648
+ intelligenceIndex: 32.5,
649
+ normalizedScore: 46,
650
+
651
+ // AA specific benchmarks
652
+ codingIndex: 29.5,
653
+ mathIndex: 86.0,
654
+
655
+ // Academic benchmarks
656
+ mmluPro: 0.829,
657
+ gpqa: 0.78,
658
+ hle: 0.133,
659
+
660
+ // Capabilities
661
+ contextWindow: 8192,
662
+ supportsReasoning: false,
663
+ supportsVision: false,
664
+
665
+ // Metadata
666
+ lastUpdated: "2026-04-06",
667
+ },
668
+ "glm-4.7-flash-reasoning": {
669
+ // AA Intelligence Index (composite score)
670
+ intelligenceIndex: 30.1,
671
+ normalizedScore: 43,
672
+
673
+ // AA specific benchmarks
674
+ codingIndex: 25.9,
675
+ mathIndex: undefined,
676
+
677
+ // Academic benchmarks
678
+ mmluPro: undefined,
679
+ gpqa: 0.581,
680
+ hle: 0.071,
681
+
682
+ // Capabilities
683
+ contextWindow: 8192,
684
+ supportsReasoning: false,
685
+ supportsVision: false,
686
+
687
+ // Metadata
688
+ lastUpdated: "2026-04-06",
689
+ },
690
+ "glm-4.5v-reasoning": {
691
+ // AA Intelligence Index (composite score)
692
+ intelligenceIndex: 15.1,
693
+ normalizedScore: 22,
694
+
695
+ // AA specific benchmarks
696
+ codingIndex: 10.9,
697
+ mathIndex: 73.0,
698
+
699
+ // Academic benchmarks
700
+ mmluPro: 0.788,
701
+ gpqa: 0.684,
702
+ hle: 0.059,
703
+
704
+ // Capabilities
705
+ contextWindow: 8192,
706
+ supportsReasoning: false,
707
+ supportsVision: false,
708
+
709
+ // Metadata
710
+ lastUpdated: "2026-04-06",
711
+ },
712
+ "glm-4.5-air": {
713
+ // AA Intelligence Index (composite score)
714
+ intelligenceIndex: 23.2,
715
+ normalizedScore: 33,
716
+
717
+ // AA specific benchmarks
718
+ codingIndex: 23.8,
719
+ mathIndex: 80.7,
720
+
721
+ // Academic benchmarks
722
+ mmluPro: 0.815,
723
+ gpqa: 0.733,
724
+ hle: 0.068,
725
+
726
+ // Capabilities
727
+ contextWindow: 8192,
728
+ supportsReasoning: false,
729
+ supportsVision: false,
730
+
731
+ // Metadata
732
+ lastUpdated: "2026-04-06",
733
+ },
734
+ "command-r-apr-24": {
735
+ // AA Intelligence Index (composite score)
736
+ intelligenceIndex: 8.3,
737
+ normalizedScore: 12,
738
+
739
+ // AA specific benchmarks
740
+ codingIndex: undefined,
741
+ mathIndex: undefined,
742
+
743
+ // Academic benchmarks
744
+ mmluPro: 0.432,
745
+ gpqa: 0.323,
746
+ hle: 0.045,
747
+
748
+ // Capabilities
749
+ contextWindow: 8192,
750
+ supportsReasoning: false,
751
+ supportsVision: false,
752
+
753
+ // Metadata
754
+ lastUpdated: "2026-04-06",
755
+ },
756
+ "command-r-mar-24": {
757
+ // AA Intelligence Index (composite score)
758
+ intelligenceIndex: 7.4,
759
+ normalizedScore: 11,
760
+
761
+ // AA specific benchmarks
762
+ codingIndex: undefined,
763
+ mathIndex: undefined,
764
+
765
+ // Academic benchmarks
766
+ mmluPro: 0.338,
767
+ gpqa: 0.284,
768
+ hle: 0.048,
769
+
770
+ // Capabilities
771
+ contextWindow: 8192,
772
+ supportsReasoning: false,
773
+ supportsVision: false,
774
+
775
+ // Metadata
776
+ lastUpdated: "2026-04-06",
777
+ },
778
+ "apriel-v1.5-15b-thinker": {
779
+ // AA Intelligence Index (composite score)
780
+ intelligenceIndex: 28.3,
781
+ normalizedScore: 40,
782
+
783
+ // AA specific benchmarks
784
+ codingIndex: 18.7,
785
+ mathIndex: 87.5,
786
+
787
+ // Academic benchmarks
788
+ mmluPro: 0.773,
789
+ gpqa: 0.713,
790
+ hle: 0.12,
791
+
792
+ // Capabilities
793
+ contextWindow: 8192,
794
+ supportsReasoning: false,
795
+ supportsVision: false,
796
+
797
+ // Metadata
798
+ lastUpdated: "2026-04-06",
799
+ },
800
+ "jamba-1.6-large": {
801
+ // AA Intelligence Index (composite score)
802
+ intelligenceIndex: 10.6,
803
+ normalizedScore: 15,
804
+
805
+ // AA specific benchmarks
806
+ codingIndex: undefined,
807
+ mathIndex: undefined,
808
+
809
+ // Academic benchmarks
810
+ mmluPro: 0.565,
811
+ gpqa: 0.387,
812
+ hle: 0.04,
813
+
814
+ // Capabilities
815
+ contextWindow: 8192,
816
+ supportsReasoning: false,
817
+ supportsVision: false,
818
+
819
+ // Metadata
820
+ lastUpdated: "2026-04-06",
821
+ },
822
+ "jamba-1.5-large": {
823
+ // AA Intelligence Index (composite score)
824
+ intelligenceIndex: 10.7,
825
+ normalizedScore: 15,
826
+
827
+ // AA specific benchmarks
828
+ codingIndex: undefined,
829
+ mathIndex: undefined,
830
+
831
+ // Academic benchmarks
832
+ mmluPro: 0.572,
833
+ gpqa: 0.427,
834
+ hle: 0.04,
835
+
836
+ // Capabilities
837
+ contextWindow: 8192,
838
+ supportsReasoning: false,
839
+ supportsVision: false,
840
+
841
+ // Metadata
842
+ lastUpdated: "2026-04-06",
843
+ },
844
+ "jamba-1.5-mini": {
845
+ // AA Intelligence Index (composite score)
846
+ intelligenceIndex: 8.0,
847
+ normalizedScore: 11,
848
+
849
+ // AA specific benchmarks
850
+ codingIndex: undefined,
851
+ mathIndex: undefined,
852
+
853
+ // Academic benchmarks
854
+ mmluPro: 0.371,
855
+ gpqa: 0.302,
856
+ hle: 0.051,
857
+
858
+ // Capabilities
859
+ contextWindow: 8192,
860
+ supportsReasoning: false,
861
+ supportsVision: false,
862
+
863
+ // Metadata
864
+ lastUpdated: "2026-04-06",
865
+ },
866
+ "jamba-1.6-mini": {
867
+ // AA Intelligence Index (composite score)
868
+ intelligenceIndex: 7.9,
869
+ normalizedScore: 11,
870
+
871
+ // AA specific benchmarks
872
+ codingIndex: undefined,
873
+ mathIndex: undefined,
874
+
875
+ // Academic benchmarks
876
+ mmluPro: 0.367,
877
+ gpqa: 0.3,
878
+ hle: 0.046,
879
+
880
+ // Capabilities
881
+ contextWindow: 8192,
882
+ supportsReasoning: false,
883
+ supportsVision: false,
884
+
885
+ // Metadata
886
+ lastUpdated: "2026-04-06",
887
+ },
888
+ "arctic-instruct": {
889
+ // AA Intelligence Index (composite score)
890
+ intelligenceIndex: 8.8,
891
+ normalizedScore: 13,
892
+
893
+ // AA specific benchmarks
894
+ codingIndex: undefined,
895
+ mathIndex: undefined,
896
+
897
+ // Academic benchmarks
898
+ mmluPro: undefined,
899
+ gpqa: undefined,
900
+ hle: undefined,
901
+
902
+ // Capabilities
903
+ contextWindow: 8192,
904
+ supportsReasoning: false,
905
+ supportsVision: false,
906
+
907
+ // Metadata
908
+ lastUpdated: "2026-04-06",
909
+ },
910
+ "qwen2.5-max": {
911
+ // AA Intelligence Index (composite score)
912
+ intelligenceIndex: 16.3,
913
+ normalizedScore: 23,
914
+
915
+ // AA specific benchmarks
916
+ codingIndex: undefined,
917
+ mathIndex: undefined,
918
+
919
+ // Academic benchmarks
920
+ mmluPro: 0.762,
921
+ gpqa: 0.587,
922
+ hle: 0.045,
923
+
924
+ // Capabilities
925
+ contextWindow: 8192,
926
+ supportsReasoning: false,
927
+ supportsVision: false,
928
+
929
+ // Metadata
930
+ lastUpdated: "2026-04-06",
931
+ },
932
+ "qwen2.5-instruct-72b": {
933
+ // AA Intelligence Index (composite score)
934
+ intelligenceIndex: 15.6,
935
+ normalizedScore: 22,
936
+
937
+ // AA specific benchmarks
938
+ codingIndex: 11.9,
939
+ mathIndex: 14.0,
940
+
941
+ // Academic benchmarks
942
+ mmluPro: 0.72,
943
+ gpqa: 0.491,
944
+ hle: 0.042,
945
+
946
+ // Capabilities
947
+ contextWindow: 8192,
948
+ supportsReasoning: false,
949
+ supportsVision: false,
950
+
951
+ // Metadata
952
+ lastUpdated: "2026-04-06",
953
+ },
954
+ "qwen2.5-coder-instruct-32b": {
955
+ // AA Intelligence Index (composite score)
956
+ intelligenceIndex: 12.9,
957
+ normalizedScore: 18,
958
+
959
+ // AA specific benchmarks
960
+ codingIndex: undefined,
961
+ mathIndex: undefined,
962
+
963
+ // Academic benchmarks
964
+ mmluPro: 0.635,
965
+ gpqa: 0.417,
966
+ hle: 0.038,
967
+
968
+ // Capabilities
969
+ contextWindow: 8192,
970
+ supportsReasoning: false,
971
+ supportsVision: false,
972
+
973
+ // Metadata
974
+ lastUpdated: "2026-04-06",
975
+ },
976
+ "qwen2.5-turbo": {
977
+ // AA Intelligence Index (composite score)
978
+ intelligenceIndex: 12.0,
979
+ normalizedScore: 17,
980
+
981
+ // AA specific benchmarks
982
+ codingIndex: undefined,
983
+ mathIndex: undefined,
984
+
985
+ // Academic benchmarks
986
+ mmluPro: 0.633,
987
+ gpqa: 0.41,
988
+ hle: 0.042,
989
+
990
+ // Capabilities
991
+ contextWindow: 8192,
992
+ supportsReasoning: false,
993
+ supportsVision: false,
994
+
995
+ // Metadata
996
+ lastUpdated: "2026-04-06",
997
+ },
998
+ "qwen2-instruct-72b": {
999
+ // AA Intelligence Index (composite score)
1000
+ intelligenceIndex: 11.7,
1001
+ normalizedScore: 17,
1002
+
1003
+ // AA specific benchmarks
1004
+ codingIndex: undefined,
1005
+ mathIndex: undefined,
1006
+
1007
+ // Academic benchmarks
1008
+ mmluPro: 0.622,
1009
+ gpqa: 0.371,
1010
+ hle: 0.037,
1011
+
1012
+ // Capabilities
1013
+ contextWindow: 8192,
1014
+ supportsReasoning: false,
1015
+ supportsVision: false,
1016
+
1017
+ // Metadata
1018
+ lastUpdated: "2026-04-06",
1019
+ },
1020
+ "qwen3-vl-30b-a3b-instruct": {
1021
+ // AA Intelligence Index (composite score)
1022
+ intelligenceIndex: 16.1,
1023
+ normalizedScore: 23,
1024
+
1025
+ // AA specific benchmarks
1026
+ codingIndex: 14.3,
1027
+ mathIndex: 72.3,
1028
+
1029
+ // Academic benchmarks
1030
+ mmluPro: 0.764,
1031
+ gpqa: 0.695,
1032
+ hle: 0.064,
1033
+
1034
+ // Capabilities
1035
+ contextWindow: 8192,
1036
+ supportsReasoning: false,
1037
+ supportsVision: false,
1038
+
1039
+ // Metadata
1040
+ lastUpdated: "2026-04-06",
1041
+ },
1042
+ "qwen3-235b-a22b-2507-instruct": {
1043
+ // AA Intelligence Index (composite score)
1044
+ intelligenceIndex: 25.0,
1045
+ normalizedScore: 36,
1046
+
1047
+ // AA specific benchmarks
1048
+ codingIndex: 22.1,
1049
+ mathIndex: 71.7,
1050
+
1051
+ // Academic benchmarks
1052
+ mmluPro: 0.828,
1053
+ gpqa: 0.753,
1054
+ hle: 0.106,
1055
+
1056
+ // Capabilities
1057
+ contextWindow: 8192,
1058
+ supportsReasoning: false,
1059
+ supportsVision: false,
1060
+
1061
+ // Metadata
1062
+ lastUpdated: "2026-04-06",
1063
+ },
1064
+ "qwen3-32b-non-reasoning": {
1065
+ // AA Intelligence Index (composite score)
1066
+ intelligenceIndex: 14.5,
1067
+ normalizedScore: 21,
1068
+
1069
+ // AA specific benchmarks
1070
+ codingIndex: undefined,
1071
+ mathIndex: 19.7,
1072
+
1073
+ // Academic benchmarks
1074
+ mmluPro: 0.727,
1075
+ gpqa: 0.535,
1076
+ hle: 0.043,
1077
+
1078
+ // Capabilities
1079
+ contextWindow: 8192,
1080
+ supportsReasoning: false,
1081
+ supportsVision: false,
1082
+
1083
+ // Metadata
1084
+ lastUpdated: "2026-04-06",
1085
+ },
1086
+ "qwen3-235b-a22b-reasoning": {
1087
+ // AA Intelligence Index (composite score)
1088
+ intelligenceIndex: 19.8,
1089
+ normalizedScore: 28,
1090
+
1091
+ // AA specific benchmarks
1092
+ codingIndex: 17.4,
1093
+ mathIndex: 82.0,
1094
+
1095
+ // Academic benchmarks
1096
+ mmluPro: 0.828,
1097
+ gpqa: 0.7,
1098
+ hle: 0.117,
1099
+
1100
+ // Capabilities
1101
+ contextWindow: 8192,
1102
+ supportsReasoning: false,
1103
+ supportsVision: false,
1104
+
1105
+ // Metadata
1106
+ lastUpdated: "2026-04-06",
1107
+ },
1108
+ "qwen3-235b-a22b-non-reasoning": {
1109
+ // AA Intelligence Index (composite score)
1110
+ intelligenceIndex: 17.0,
1111
+ normalizedScore: 24,
1112
+
1113
+ // AA specific benchmarks
1114
+ codingIndex: 14.0,
1115
+ mathIndex: 23.7,
1116
+
1117
+ // Academic benchmarks
1118
+ mmluPro: 0.762,
1119
+ gpqa: 0.613,
1120
+ hle: 0.047,
1121
+
1122
+ // Capabilities
1123
+ contextWindow: 8192,
1124
+ supportsReasoning: false,
1125
+ supportsVision: false,
1126
+
1127
+ // Metadata
1128
+ lastUpdated: "2026-04-06",
1129
+ },
1130
+ "qwen3-30b-a3b-2507-reasoning": {
1131
+ // AA Intelligence Index (composite score)
1132
+ intelligenceIndex: 22.4,
1133
+ normalizedScore: 32,
1134
+
1135
+ // AA specific benchmarks
1136
+ codingIndex: 14.7,
1137
+ mathIndex: 56.3,
1138
+
1139
+ // Academic benchmarks
1140
+ mmluPro: 0.805,
1141
+ gpqa: 0.707,
1142
+ hle: 0.098,
1143
+
1144
+ // Capabilities
1145
+ contextWindow: 8192,
1146
+ supportsReasoning: false,
1147
+ supportsVision: false,
1148
+
1149
+ // Metadata
1150
+ lastUpdated: "2026-04-06",
1151
+ },
1152
+ "qwen3-vl-235b-a22b-instruct": {
1153
+ // AA Intelligence Index (composite score)
1154
+ intelligenceIndex: 20.8,
1155
+ normalizedScore: 30,
1156
+
1157
+ // AA specific benchmarks
1158
+ codingIndex: 16.5,
1159
+ mathIndex: 70.7,
1160
+
1161
+ // Academic benchmarks
1162
+ mmluPro: 0.823,
1163
+ gpqa: 0.712,
1164
+ hle: 0.063,
1165
+
1166
+ // Capabilities
1167
+ contextWindow: 8192,
1168
+ supportsReasoning: false,
1169
+ supportsVision: false,
1170
+
1171
+ // Metadata
1172
+ lastUpdated: "2026-04-06",
1173
+ },
1174
+ "qwen3-0.6b-reasoning": {
1175
+ // AA Intelligence Index (composite score)
1176
+ intelligenceIndex: 6.5,
1177
+ normalizedScore: 9,
1178
+
1179
+ // AA specific benchmarks
1180
+ codingIndex: 0.9,
1181
+ mathIndex: 18.0,
1182
+
1183
+ // Academic benchmarks
1184
+ mmluPro: 0.347,
1185
+ gpqa: 0.239,
1186
+ hle: 0.057,
1187
+
1188
+ // Capabilities
1189
+ contextWindow: 8192,
1190
+ supportsReasoning: false,
1191
+ supportsVision: false,
1192
+
1193
+ // Metadata
1194
+ lastUpdated: "2026-04-06",
1195
+ },
1196
+ "qwen3-235b-a22b-2507-reasoning": {
1197
+ // AA Intelligence Index (composite score)
1198
+ intelligenceIndex: 29.5,
1199
+ normalizedScore: 42,
1200
+
1201
+ // AA specific benchmarks
1202
+ codingIndex: 23.2,
1203
+ mathIndex: 91.0,
1204
+
1205
+ // Academic benchmarks
1206
+ mmluPro: 0.843,
1207
+ gpqa: 0.79,
1208
+ hle: 0.15,
1209
+
1210
+ // Capabilities
1211
+ contextWindow: 8192,
1212
+ supportsReasoning: false,
1213
+ supportsVision: false,
1214
+
1215
+ // Metadata
1216
+ lastUpdated: "2026-04-06",
1217
+ },
1218
+ "qwen3-8b-non-reasoning": {
1219
+ // AA Intelligence Index (composite score)
1220
+ intelligenceIndex: 10.6,
1221
+ normalizedScore: 15,
1222
+
1223
+ // AA specific benchmarks
1224
+ codingIndex: 7.1,
1225
+ mathIndex: 24.3,
1226
+
1227
+ // Academic benchmarks
1228
+ mmluPro: 0.643,
1229
+ gpqa: 0.452,
1230
+ hle: 0.028,
1231
+
1232
+ // Capabilities
1233
+ contextWindow: 8192,
1234
+ supportsReasoning: false,
1235
+ supportsVision: false,
1236
+
1237
+ // Metadata
1238
+ lastUpdated: "2026-04-06",
1239
+ },
1240
+ "qwen3-4b-2507-reasoning": {
1241
+ // AA Intelligence Index (composite score)
1242
+ intelligenceIndex: 18.2,
1243
+ normalizedScore: 26,
1244
+
1245
+ // AA specific benchmarks
1246
+ codingIndex: 9.5,
1247
+ mathIndex: 82.7,
1248
+
1249
+ // Academic benchmarks
1250
+ mmluPro: 0.743,
1251
+ gpqa: 0.667,
1252
+ hle: 0.059,
1253
+
1254
+ // Capabilities
1255
+ contextWindow: 8192,
1256
+ supportsReasoning: false,
1257
+ supportsVision: false,
1258
+
1259
+ // Metadata
1260
+ lastUpdated: "2026-04-06",
1261
+ },
1262
+ "qwq-32b": {
1263
+ // AA Intelligence Index (composite score)
1264
+ intelligenceIndex: 19.7,
1265
+ normalizedScore: 28,
1266
+
1267
+ // AA specific benchmarks
1268
+ codingIndex: undefined,
1269
+ mathIndex: 29.0,
1270
+
1271
+ // Academic benchmarks
1272
+ mmluPro: 0.764,
1273
+ gpqa: 0.593,
1274
+ hle: 0.082,
1275
+
1276
+ // Capabilities
1277
+ contextWindow: 8192,
1278
+ supportsReasoning: false,
1279
+ supportsVision: false,
1280
+
1281
+ // Metadata
1282
+ lastUpdated: "2026-04-06",
1283
+ },
1284
+ "qwen3-32b-reasoning": {
1285
+ // AA Intelligence Index (composite score)
1286
+ intelligenceIndex: 16.5,
1287
+ normalizedScore: 24,
1288
+
1289
+ // AA specific benchmarks
1290
+ codingIndex: 13.8,
1291
+ mathIndex: 73.0,
1292
+
1293
+ // Academic benchmarks
1294
+ mmluPro: 0.798,
1295
+ gpqa: 0.668,
1296
+ hle: 0.083,
1297
+
1298
+ // Capabilities
1299
+ contextWindow: 8192,
1300
+ supportsReasoning: false,
1301
+ supportsVision: false,
1302
+
1303
+ // Metadata
1304
+ lastUpdated: "2026-04-06",
1305
+ },
1306
+ "qwen3-4b-2507-instruct": {
1307
+ // AA Intelligence Index (composite score)
1308
+ intelligenceIndex: 12.9,
1309
+ normalizedScore: 18,
1310
+
1311
+ // AA specific benchmarks
1312
+ codingIndex: 9.1,
1313
+ mathIndex: 52.3,
1314
+
1315
+ // Academic benchmarks
1316
+ mmluPro: 0.672,
1317
+ gpqa: 0.517,
1318
+ hle: 0.047,
1319
+
1320
+ // Capabilities
1321
+ contextWindow: 8192,
1322
+ supportsReasoning: false,
1323
+ supportsVision: false,
1324
+
1325
+ // Metadata
1326
+ lastUpdated: "2026-04-06",
1327
+ },
1328
+ "qwen3-30b-a3b-2507-instruct": {
1329
+ // AA Intelligence Index (composite score)
1330
+ intelligenceIndex: 15.0,
1331
+ normalizedScore: 21,
1332
+
1333
+ // AA specific benchmarks
1334
+ codingIndex: 14.2,
1335
+ mathIndex: 66.3,
1336
+
1337
+ // Academic benchmarks
1338
+ mmluPro: 0.777,
1339
+ gpqa: 0.659,
1340
+ hle: 0.068,
1341
+
1342
+ // Capabilities
1343
+ contextWindow: 8192,
1344
+ supportsReasoning: false,
1345
+ supportsVision: false,
1346
+
1347
+ // Metadata
1348
+ lastUpdated: "2026-04-06",
1349
+ },
1350
+ "qwen3-14b-non-reasoning": {
1351
+ // AA Intelligence Index (composite score)
1352
+ intelligenceIndex: 12.8,
1353
+ normalizedScore: 18,
1354
+
1355
+ // AA specific benchmarks
1356
+ codingIndex: 12.4,
1357
+ mathIndex: 58.0,
1358
+
1359
+ // Academic benchmarks
1360
+ mmluPro: 0.675,
1361
+ gpqa: 0.47,
1362
+ hle: 0.042,
1363
+
1364
+ // Capabilities
1365
+ contextWindow: 8192,
1366
+ supportsReasoning: false,
1367
+ supportsVision: false,
1368
+
1369
+ // Metadata
1370
+ lastUpdated: "2026-04-06",
1371
+ },
1372
+ "qwen3-vl-4b-instruct": {
1373
+ // AA Intelligence Index (composite score)
1374
+ intelligenceIndex: 9.6,
1375
+ normalizedScore: 14,
1376
+
1377
+ // AA specific benchmarks
1378
+ codingIndex: 4.5,
1379
+ mathIndex: 37.0,
1380
+
1381
+ // Academic benchmarks
1382
+ mmluPro: 0.634,
1383
+ gpqa: 0.371,
1384
+ hle: 0.037,
1385
+
1386
+ // Capabilities
1387
+ contextWindow: 8192,
1388
+ supportsReasoning: false,
1389
+ supportsVision: false,
1390
+
1391
+ // Metadata
1392
+ lastUpdated: "2026-04-06",
1393
+ },
1394
+ "qwen2.5-coder-instruct-7b": {
1395
+ // AA Intelligence Index (composite score)
1396
+ intelligenceIndex: 10.0,
1397
+ normalizedScore: 14,
1398
+
1399
+ // AA specific benchmarks
1400
+ codingIndex: undefined,
1401
+ mathIndex: undefined,
1402
+
1403
+ // Academic benchmarks
1404
+ mmluPro: 0.473,
1405
+ gpqa: 0.339,
1406
+ hle: 0.048,
1407
+
1408
+ // Capabilities
1409
+ contextWindow: 8192,
1410
+ supportsReasoning: false,
1411
+ supportsVision: false,
1412
+
1413
+ // Metadata
1414
+ lastUpdated: "2026-04-06",
1415
+ },
1416
+ "qwen3-30b-a3b-reasoning": {
1417
+ // AA Intelligence Index (composite score)
1418
+ intelligenceIndex: 15.3,
1419
+ normalizedScore: 22,
1420
+
1421
+ // AA specific benchmarks
1422
+ codingIndex: 11.0,
1423
+ mathIndex: 72.3,
1424
+
1425
+ // Academic benchmarks
1426
+ mmluPro: 0.777,
1427
+ gpqa: 0.616,
1428
+ hle: 0.066,
1429
+
1430
+ // Capabilities
1431
+ contextWindow: 8192,
1432
+ supportsReasoning: false,
1433
+ supportsVision: false,
1434
+
1435
+ // Metadata
1436
+ lastUpdated: "2026-04-06",
1437
+ },
1438
+ "qwq-32b-preview": {
1439
+ // AA Intelligence Index (composite score)
1440
+ intelligenceIndex: 15.2,
1441
+ normalizedScore: 22,
1442
+
1443
+ // AA specific benchmarks
1444
+ codingIndex: undefined,
1445
+ mathIndex: undefined,
1446
+
1447
+ // Academic benchmarks
1448
+ mmluPro: 0.648,
1449
+ gpqa: 0.557,
1450
+ hle: 0.048,
1451
+
1452
+ // Capabilities
1453
+ contextWindow: 8192,
1454
+ supportsReasoning: false,
1455
+ supportsVision: false,
1456
+
1457
+ // Metadata
1458
+ lastUpdated: "2026-04-06",
1459
+ },
1460
+ "qwen3-vl-32b-instruct": {
1461
+ // AA Intelligence Index (composite score)
1462
+ intelligenceIndex: 17.2,
1463
+ normalizedScore: 25,
1464
+
1465
+ // AA specific benchmarks
1466
+ codingIndex: 15.6,
1467
+ mathIndex: 68.3,
1468
+
1469
+ // Academic benchmarks
1470
+ mmluPro: 0.791,
1471
+ gpqa: 0.671,
1472
+ hle: 0.063,
1473
+
1474
+ // Capabilities
1475
+ contextWindow: 8192,
1476
+ supportsReasoning: false,
1477
+ supportsVision: false,
1478
+
1479
+ // Metadata
1480
+ lastUpdated: "2026-04-06",
1481
+ },
1482
+ "qwen3-1.7b-non-reasoning": {
1483
+ // AA Intelligence Index (composite score)
1484
+ intelligenceIndex: 6.8,
1485
+ normalizedScore: 10,
1486
+
1487
+ // AA specific benchmarks
1488
+ codingIndex: 2.3,
1489
+ mathIndex: 7.3,
1490
+
1491
+ // Academic benchmarks
1492
+ mmluPro: 0.411,
1493
+ gpqa: 0.283,
1494
+ hle: 0.052,
1495
+
1496
+ // Capabilities
1497
+ contextWindow: 8192,
1498
+ supportsReasoning: false,
1499
+ supportsVision: false,
1500
+
1501
+ // Metadata
1502
+ lastUpdated: "2026-04-06",
1503
+ },
1504
+ "qwen2.5-instruct-32b": {
1505
+ // AA Intelligence Index (composite score)
1506
+ intelligenceIndex: 13.2,
1507
+ normalizedScore: 19,
1508
+
1509
+ // AA specific benchmarks
1510
+ codingIndex: undefined,
1511
+ mathIndex: undefined,
1512
+
1513
+ // Academic benchmarks
1514
+ mmluPro: 0.697,
1515
+ gpqa: 0.466,
1516
+ hle: 0.038,
1517
+
1518
+ // Capabilities
1519
+ contextWindow: 8192,
1520
+ supportsReasoning: false,
1521
+ supportsVision: false,
1522
+
1523
+ // Metadata
1524
+ lastUpdated: "2026-04-06",
1525
+ },
1526
+ "qwen3-4b-non-reasoning": {
1527
+ // AA Intelligence Index (composite score)
1528
+ intelligenceIndex: 12.5,
1529
+ normalizedScore: 18,
1530
+
1531
+ // AA specific benchmarks
1532
+ codingIndex: undefined,
1533
+ mathIndex: undefined,
1534
+
1535
+ // Academic benchmarks
1536
+ mmluPro: 0.586,
1537
+ gpqa: 0.398,
1538
+ hle: 0.037,
1539
+
1540
+ // Capabilities
1541
+ contextWindow: 8192,
1542
+ supportsReasoning: false,
1543
+ supportsVision: false,
1544
+
1545
+ // Metadata
1546
+ lastUpdated: "2026-04-06",
1547
+ },
1548
+ "qwen3-vl-30b-a3b-reasoning": {
1549
+ // AA Intelligence Index (composite score)
1550
+ intelligenceIndex: 19.7,
1551
+ normalizedScore: 28,
1552
+
1553
+ // AA specific benchmarks
1554
+ codingIndex: 13.1,
1555
+ mathIndex: 82.3,
1556
+
1557
+ // Academic benchmarks
1558
+ mmluPro: 0.807,
1559
+ gpqa: 0.72,
1560
+ hle: 0.087,
1561
+
1562
+ // Capabilities
1563
+ contextWindow: 8192,
1564
+ supportsReasoning: false,
1565
+ supportsVision: false,
1566
+
1567
+ // Metadata
1568
+ lastUpdated: "2026-04-06",
1569
+ },
1570
+ "qwen3-vl-8b-reasoning": {
1571
+ // AA Intelligence Index (composite score)
1572
+ intelligenceIndex: 16.7,
1573
+ normalizedScore: 24,
1574
+
1575
+ // AA specific benchmarks
1576
+ codingIndex: 9.8,
1577
+ mathIndex: 30.7,
1578
+
1579
+ // Academic benchmarks
1580
+ mmluPro: 0.749,
1581
+ gpqa: 0.579,
1582
+ hle: 0.033,
1583
+
1584
+ // Capabilities
1585
+ contextWindow: 8192,
1586
+ supportsReasoning: false,
1587
+ supportsVision: false,
1588
+
1589
+ // Metadata
1590
+ lastUpdated: "2026-04-06",
1591
+ },
1592
+ "qwen3-8b-reasoning": {
1593
+ // AA Intelligence Index (composite score)
1594
+ intelligenceIndex: 13.2,
1595
+ normalizedScore: 19,
1596
+
1597
+ // AA specific benchmarks
1598
+ codingIndex: 9.0,
1599
+ mathIndex: 19.0,
1600
+
1601
+ // Academic benchmarks
1602
+ mmluPro: 0.743,
1603
+ gpqa: 0.589,
1604
+ hle: 0.042,
1605
+
1606
+ // Capabilities
1607
+ contextWindow: 8192,
1608
+ supportsReasoning: false,
1609
+ supportsVision: false,
1610
+
1611
+ // Metadata
1612
+ lastUpdated: "2026-04-06",
1613
+ },
1614
+ "qwen3-vl-235b-a22b-reasoning": {
1615
+ // AA Intelligence Index (composite score)
1616
+ intelligenceIndex: 27.6,
1617
+ normalizedScore: 39,
1618
+
1619
+ // AA specific benchmarks
1620
+ codingIndex: 20.9,
1621
+ mathIndex: 88.3,
1622
+
1623
+ // Academic benchmarks
1624
+ mmluPro: 0.836,
1625
+ gpqa: 0.772,
1626
+ hle: 0.101,
1627
+
1628
+ // Capabilities
1629
+ contextWindow: 8192,
1630
+ supportsReasoning: false,
1631
+ supportsVision: false,
1632
+
1633
+ // Metadata
1634
+ lastUpdated: "2026-04-06",
1635
+ },
1636
+ "qwen3-4b-reasoning": {
1637
+ // AA Intelligence Index (composite score)
1638
+ intelligenceIndex: 14.2,
1639
+ normalizedScore: 20,
1640
+
1641
+ // AA specific benchmarks
1642
+ codingIndex: undefined,
1643
+ mathIndex: 22.3,
1644
+
1645
+ // Academic benchmarks
1646
+ mmluPro: 0.696,
1647
+ gpqa: 0.522,
1648
+ hle: 0.051,
1649
+
1650
+ // Capabilities
1651
+ contextWindow: 8192,
1652
+ supportsReasoning: false,
1653
+ supportsVision: false,
1654
+
1655
+ // Metadata
1656
+ lastUpdated: "2026-04-06",
1657
+ },
1658
+ "qwen3-0.6b-non-reasoning": {
1659
+ // AA Intelligence Index (composite score)
1660
+ intelligenceIndex: 5.7,
1661
+ normalizedScore: 8,
1662
+
1663
+ // AA specific benchmarks
1664
+ codingIndex: 1.4,
1665
+ mathIndex: 10.3,
1666
+
1667
+ // Academic benchmarks
1668
+ mmluPro: 0.231,
1669
+ gpqa: 0.231,
1670
+ hle: 0.052,
1671
+
1672
+ // Capabilities
1673
+ contextWindow: 8192,
1674
+ supportsReasoning: false,
1675
+ supportsVision: false,
1676
+
1677
+ // Metadata
1678
+ lastUpdated: "2026-04-06",
1679
+ },
1680
+ "qwen3-30b-a3b-non-reasoning": {
1681
+ // AA Intelligence Index (composite score)
1682
+ intelligenceIndex: 12.5,
1683
+ normalizedScore: 18,
1684
+
1685
+ // AA specific benchmarks
1686
+ codingIndex: 13.3,
1687
+ mathIndex: 21.7,
1688
+
1689
+ // Academic benchmarks
1690
+ mmluPro: 0.71,
1691
+ gpqa: 0.515,
1692
+ hle: 0.046,
1693
+
1694
+ // Capabilities
1695
+ contextWindow: 8192,
1696
+ supportsReasoning: false,
1697
+ supportsVision: false,
1698
+
1699
+ // Metadata
1700
+ lastUpdated: "2026-04-06",
1701
+ },
1702
+ "qwen3-14b-reasoning": {
1703
+ // AA Intelligence Index (composite score)
1704
+ intelligenceIndex: 16.2,
1705
+ normalizedScore: 23,
1706
+
1707
+ // AA specific benchmarks
1708
+ codingIndex: 13.1,
1709
+ mathIndex: 55.7,
1710
+
1711
+ // Academic benchmarks
1712
+ mmluPro: 0.774,
1713
+ gpqa: 0.604,
1714
+ hle: 0.043,
1715
+
1716
+ // Capabilities
1717
+ contextWindow: 8192,
1718
+ supportsReasoning: false,
1719
+ supportsVision: false,
1720
+
1721
+ // Metadata
1722
+ lastUpdated: "2026-04-06",
1723
+ },
1724
+ "qwen3-1.7b-reasoning": {
1725
+ // AA Intelligence Index (composite score)
1726
+ intelligenceIndex: 8.0,
1727
+ normalizedScore: 11,
1728
+
1729
+ // AA specific benchmarks
1730
+ codingIndex: 1.4,
1731
+ mathIndex: 38.7,
1732
+
1733
+ // Academic benchmarks
1734
+ mmluPro: 0.57,
1735
+ gpqa: 0.356,
1736
+ hle: 0.048,
1737
+
1738
+ // Capabilities
1739
+ contextWindow: 8192,
1740
+ supportsReasoning: false,
1741
+ supportsVision: false,
1742
+
1743
+ // Metadata
1744
+ lastUpdated: "2026-04-06",
1745
+ },
1746
+ "qwen3-max": {
1747
+ // AA Intelligence Index (composite score)
1748
+ intelligenceIndex: 31.4,
1749
+ normalizedScore: 45,
1750
+
1751
+ // AA specific benchmarks
1752
+ codingIndex: 26.4,
1753
+ mathIndex: 80.7,
1754
+
1755
+ // Academic benchmarks
1756
+ mmluPro: 0.841,
1757
+ gpqa: 0.764,
1758
+ hle: 0.111,
1759
+
1760
+ // Capabilities
1761
+ contextWindow: 8192,
1762
+ supportsReasoning: false,
1763
+ supportsVision: false,
1764
+
1765
+ // Metadata
1766
+ lastUpdated: "2026-04-06",
1767
+ },
1768
+ "qwen3-coder-30b-a3b-instruct": {
1769
+ // AA Intelligence Index (composite score)
1770
+ intelligenceIndex: 20.0,
1771
+ normalizedScore: 29,
1772
+
1773
+ // AA specific benchmarks
1774
+ codingIndex: 19.4,
1775
+ mathIndex: 29.0,
1776
+
1777
+ // Academic benchmarks
1778
+ mmluPro: 0.706,
1779
+ gpqa: 0.516,
1780
+ hle: 0.04,
1781
+
1782
+ // Capabilities
1783
+ contextWindow: 8192,
1784
+ supportsReasoning: false,
1785
+ supportsVision: false,
1786
+
1787
+ // Metadata
1788
+ lastUpdated: "2026-04-06",
1789
+ },
1790
+ "qwen3-vl-8b-instruct": {
1791
+ // AA Intelligence Index (composite score)
1792
+ intelligenceIndex: 14.3,
1793
+ normalizedScore: 20,
1794
+
1795
+ // AA specific benchmarks
1796
+ codingIndex: 7.3,
1797
+ mathIndex: 27.3,
1798
+
1799
+ // Academic benchmarks
1800
+ mmluPro: 0.686,
1801
+ gpqa: 0.427,
1802
+ hle: 0.029,
1803
+
1804
+ // Capabilities
1805
+ contextWindow: 8192,
1806
+ supportsReasoning: false,
1807
+ supportsVision: false,
1808
+
1809
+ // Metadata
1810
+ lastUpdated: "2026-04-06",
1811
+ },
1812
+ "qwen1.5-chat-110b": {
1813
+ // AA Intelligence Index (composite score)
1814
+ intelligenceIndex: 9.5,
1815
+ normalizedScore: 14,
1816
+
1817
+ // AA specific benchmarks
1818
+ codingIndex: undefined,
1819
+ mathIndex: undefined,
1820
+
1821
+ // Academic benchmarks
1822
+ mmluPro: undefined,
1823
+ gpqa: 0.289,
1824
+ hle: undefined,
1825
+
1826
+ // Capabilities
1827
+ contextWindow: 8192,
1828
+ supportsReasoning: false,
1829
+ supportsVision: false,
1830
+
1831
+ // Metadata
1832
+ lastUpdated: "2026-04-06",
1833
+ },
1834
+ "qwen3-max-preview": {
1835
+ // AA Intelligence Index (composite score)
1836
+ intelligenceIndex: 26.1,
1837
+ normalizedScore: 37,
1838
+
1839
+ // AA specific benchmarks
1840
+ codingIndex: 25.5,
1841
+ mathIndex: 75.0,
1842
+
1843
+ // Academic benchmarks
1844
+ mmluPro: 0.838,
1845
+ gpqa: 0.764,
1846
+ hle: 0.093,
1847
+
1848
+ // Capabilities
1849
+ contextWindow: 8192,
1850
+ supportsReasoning: false,
1851
+ supportsVision: false,
1852
+
1853
+ // Metadata
1854
+ lastUpdated: "2026-04-06",
1855
+ },
1856
+ "qwen3-vl-4b-reasoning": {
1857
+ // AA Intelligence Index (composite score)
1858
+ intelligenceIndex: 13.7,
1859
+ normalizedScore: 20,
1860
+
1861
+ // AA specific benchmarks
1862
+ codingIndex: 6.7,
1863
+ mathIndex: 25.7,
1864
+
1865
+ // Academic benchmarks
1866
+ mmluPro: 0.7,
1867
+ gpqa: 0.494,
1868
+ hle: 0.044,
1869
+
1870
+ // Capabilities
1871
+ contextWindow: 8192,
1872
+ supportsReasoning: false,
1873
+ supportsVision: false,
1874
+
1875
+ // Metadata
1876
+ lastUpdated: "2026-04-06",
1877
+ },
1878
+ "qwen3-max-thinking-preview": {
1879
+ // AA Intelligence Index (composite score)
1880
+ intelligenceIndex: 32.5,
1881
+ normalizedScore: 46,
1882
+
1883
+ // AA specific benchmarks
1884
+ codingIndex: 24.5,
1885
+ mathIndex: 82.3,
1886
+
1887
+ // Academic benchmarks
1888
+ mmluPro: 0.824,
1889
+ gpqa: 0.776,
1890
+ hle: 0.12,
1891
+
1892
+ // Capabilities
1893
+ contextWindow: 8192,
1894
+ supportsReasoning: false,
1895
+ supportsVision: false,
1896
+
1897
+ // Metadata
1898
+ lastUpdated: "2026-04-06",
1899
+ },
1900
+ "qwen3-vl-32b-reasoning": {
1901
+ // AA Intelligence Index (composite score)
1902
+ intelligenceIndex: 24.7,
1903
+ normalizedScore: 35,
1904
+
1905
+ // AA specific benchmarks
1906
+ codingIndex: 14.5,
1907
+ mathIndex: 84.7,
1908
+
1909
+ // Academic benchmarks
1910
+ mmluPro: 0.818,
1911
+ gpqa: 0.733,
1912
+ hle: 0.096,
1913
+
1914
+ // Capabilities
1915
+ contextWindow: 8192,
1916
+ supportsReasoning: false,
1917
+ supportsVision: false,
1918
+
1919
+ // Metadata
1920
+ lastUpdated: "2026-04-06",
1921
+ },
1922
+ "qwen-chat-72b": {
1923
+ // AA Intelligence Index (composite score)
1924
+ intelligenceIndex: 8.8,
1925
+ normalizedScore: 13,
1926
+
1927
+ // AA specific benchmarks
1928
+ codingIndex: undefined,
1929
+ mathIndex: undefined,
1930
+
1931
+ // Academic benchmarks
1932
+ mmluPro: undefined,
1933
+ gpqa: undefined,
1934
+ hle: undefined,
1935
+
1936
+ // Capabilities
1937
+ contextWindow: 8192,
1938
+ supportsReasoning: false,
1939
+ supportsVision: false,
1940
+
1941
+ // Metadata
1942
+ lastUpdated: "2026-04-06",
1943
+ },
1944
+ "seed-oss-36b-instruct": {
1945
+ // AA Intelligence Index (composite score)
1946
+ intelligenceIndex: 25.2,
1947
+ normalizedScore: 36,
1948
+
1949
+ // AA specific benchmarks
1950
+ codingIndex: 16.7,
1951
+ mathIndex: 84.7,
1952
+
1953
+ // Academic benchmarks
1954
+ mmluPro: 0.815,
1955
+ gpqa: 0.726,
1956
+ hle: 0.091,
1957
+
1958
+ // Capabilities
1959
+ contextWindow: 8192,
1960
+ supportsReasoning: false,
1961
+ supportsVision: false,
1962
+
1963
+ // Metadata
1964
+ lastUpdated: "2026-04-06",
1965
+ },
1966
+
1967
+
1968
+
1969
+ };