pi-free 2.0.12 → 2.0.14

package/provider-failover/benchmarks-chunk-5.ts

@@ -0,0 +1,1301 @@
+// Auto-generated benchmark data chunk 5
+// Models: glm-4.5-air .. seed-oss-36b-instruct (68 entries)
+// Last updated: 2026-06-01
+// DO NOT EDIT MANUALLY — generated by scripts/update-benchmarks.ts
+
+import type { HardcodedBenchmark } from "./hardcoded-benchmarks.ts";
+
+export const BENCHMARKS_CHUNK_5: Record<string, HardcodedBenchmark> = {
+	"glm-4.5-air": {
+		// AA specific benchmarks
+		codingIndex: 23.8,
+		mathIndex: 80.7,
+
+		// Academic benchmarks
+		mmluPro: 0.815,
+		gpqa: 0.733,
+		hle: 0.068,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "GLM-4.5-Air",
+	},
+	"glm-4.6v-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 19.7,
+		mathIndex: 85.3,
+
+		// Academic benchmarks
+		mmluPro: 0.799,
+		gpqa: 0.719,
+		hle: 0.089,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "GLM-4.6V (Reasoning)",
+	},
+	"glm-4.5v-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 10.8,
+		mathIndex: 15.3,
+
+		// Academic benchmarks
+		mmluPro: 0.751,
+		gpqa: 0.573,
+		hle: 0.036,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "GLM-4.5V (Non-reasoning)",
+	},
+	"command-r+-apr-24": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.432,
+		gpqa: 0.323,
+		hle: 0.045,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Command-R+ (Apr '24)",
+	},
+	"command-r-mar-24": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.338,
+		gpqa: 0.284,
+		hle: 0.048,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Command-R (Mar '24)",
+	},
+	"apriel-v1.5-15b-thinker": {
+		// AA specific benchmarks
+		codingIndex: 18.7,
+		mathIndex: 87.5,
+
+		// Academic benchmarks
+		mmluPro: 0.773,
+		gpqa: 0.713,
+		hle: 0.12,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Apriel-v1.5-15B-Thinker",
+	},
+	"jamba-1.5-large": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.572,
+		gpqa: 0.427,
+		hle: 0.04,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Jamba 1.5 Large",
+	},
+	"jamba-1.5-mini": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.371,
+		gpqa: 0.302,
+		hle: 0.051,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Jamba 1.5 Mini",
+	},
+	"jamba-1.6-mini": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.367,
+		gpqa: 0.3,
+		hle: 0.046,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Jamba 1.6 Mini",
+	},
+	"jamba-1.6-large": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.565,
+		gpqa: 0.387,
+		hle: 0.04,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Jamba 1.6 Large",
+	},
+	"arctic-instruct": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: undefined,
+		hle: undefined,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Arctic Instruct",
+	},
+	"qwen2.5-max": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.762,
+		gpqa: 0.587,
+		hle: 0.045,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Max",
+	},
+	"qwen2.5-instruct-72b": {
+		// AA specific benchmarks
+		codingIndex: 11.9,
+		mathIndex: 14,
+
+		// Academic benchmarks
+		mmluPro: 0.72,
+		gpqa: 0.491,
+		hle: 0.042,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Instruct 72B",
+	},
+	"qwen2.5-coder-instruct-32b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.635,
+		gpqa: 0.417,
+		hle: 0.038,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Coder Instruct 32B",
+	},
+	"qwen2.5-turbo": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.633,
+		gpqa: 0.41,
+		hle: 0.042,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Turbo",
+	},
+	"qwen2-instruct-72b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.622,
+		gpqa: 0.371,
+		hle: 0.037,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2 Instruct 72B",
+	},
+	"qwq-32b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: 29,
+
+		// Academic benchmarks
+		mmluPro: 0.764,
+		gpqa: 0.593,
+		hle: 0.082,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "QwQ 32B",
+	},
+	"qwen3-vl-235b-a22b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 16.5,
+		mathIndex: 70.7,
+
+		// Academic benchmarks
+		mmluPro: 0.823,
+		gpqa: 0.712,
+		hle: 0.063,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 235B A22B Instruct",
+	},
+	"qwen3-coder-30b-a3b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 19.4,
+		mathIndex: 29,
+
+		// Academic benchmarks
+		mmluPro: 0.706,
+		gpqa: 0.516,
+		hle: 0.04,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Coder 30B A3B Instruct",
+	},
+	"qwen3-4b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.586,
+		gpqa: 0.398,
+		hle: 0.037,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 4B (Non-reasoning)",
+	},
+	"qwen3-235b-a22b-2507-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 23.2,
+		mathIndex: 91,
+
+		// Academic benchmarks
+		mmluPro: 0.843,
+		gpqa: 0.79,
+		hle: 0.15,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 235B A22B 2507 (Reasoning)",
+	},
+	"qwen3-235b-a22b-2507-instruct": {
+		// AA specific benchmarks
+		codingIndex: 22.1,
+		mathIndex: 71.7,
+
+		// Academic benchmarks
+		mmluPro: 0.828,
+		gpqa: 0.753,
+		hle: 0.106,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 235B A22B 2507 Instruct",
+	},
+	"qwen3-coder-480b-a35b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 24.6,
+		mathIndex: 39.3,
+
+		// Academic benchmarks
+		mmluPro: 0.788,
+		gpqa: 0.618,
+		hle: 0.044,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Coder 480B A35B Instruct",
+	},
+	"qwen3.5-27b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 34.9,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.858,
+		hle: 0.222,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3.5 27B (Reasoning)",
+	},
+	"qwen3-235b-a22b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 17.4,
+		mathIndex: 82,
+
+		// Academic benchmarks
+		mmluPro: 0.828,
+		gpqa: 0.7,
+		hle: 0.117,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 235B A22B (Reasoning)",
+	},
+	"qwen3-235b-a22b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 14,
+		mathIndex: 23.7,
+
+		// Academic benchmarks
+		mmluPro: 0.762,
+		gpqa: 0.613,
+		hle: 0.047,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 235B A22B (Non-reasoning)",
+	},
+	"qwen3.5-35b-a3b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 30.3,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.845,
+		hle: 0.197,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3.5 35B A3B (Reasoning)",
+	},
+	"qwen3-32b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: 19.7,
+
+		// Academic benchmarks
+		mmluPro: 0.727,
+		gpqa: 0.535,
+		hle: 0.043,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 32B (Non-reasoning)",
+	},
+	"qwq-32b-preview": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.648,
+		gpqa: 0.557,
+		hle: 0.048,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "QwQ 32B-Preview",
+	},
+	"qwen3-vl-4b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 6.7,
+		mathIndex: 25.7,
+
+		// Academic benchmarks
+		mmluPro: 0.7,
+		gpqa: 0.494,
+		hle: 0.044,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 4B (Reasoning)",
+	},
+	"qwen3-vl-4b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 4.6,
+		mathIndex: 37,
+
+		// Academic benchmarks
+		mmluPro: 0.634,
+		gpqa: 0.371,
+		hle: 0.037,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 4B Instruct",
+	},
+	"qwen3-30b-a3b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 13.3,
+		mathIndex: 21.7,
+
+		// Academic benchmarks
+		mmluPro: 0.71,
+		gpqa: 0.515,
+		hle: 0.046,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 30B A3B (Non-reasoning)",
+	},
+	"qwen3-14b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 12.4,
+		mathIndex: 58,
+
+		// Academic benchmarks
+		mmluPro: 0.675,
+		gpqa: 0.47,
+		hle: 0.042,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 14B (Non-reasoning)",
+	},
+	"qwen3-32b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 13.8,
+		mathIndex: 73,
+
+		// Academic benchmarks
+		mmluPro: 0.798,
+		gpqa: 0.668,
+		hle: 0.083,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 32B (Reasoning)",
+	},
+	"qwen3-vl-8b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 9.8,
+		mathIndex: 30.7,
+
+		// Academic benchmarks
+		mmluPro: 0.749,
+		gpqa: 0.579,
+		hle: 0.033,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 8B (Reasoning)",
+	},
+	"qwen3-vl-30b-a3b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 13.1,
+		mathIndex: 82.3,
+
+		// Academic benchmarks
+		mmluPro: 0.807,
+		gpqa: 0.72,
+		hle: 0.087,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 30B A3B (Reasoning)",
+	},
+	"qwen3-vl-235b-a22b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 20.9,
+		mathIndex: 88.3,
+
+		// Academic benchmarks
+		mmluPro: 0.836,
+		gpqa: 0.772,
+		hle: 0.101,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 235B A22B (Reasoning)",
+	},
+	"qwen3-8b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 7.1,
+		mathIndex: 24.3,
+
+		// Academic benchmarks
+		mmluPro: 0.643,
+		gpqa: 0.452,
+		hle: 0.028,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 8B (Non-reasoning)",
+	},
+	"qwen3-8b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 9,
+		mathIndex: 19,
+
+		// Academic benchmarks
+		mmluPro: 0.743,
+		gpqa: 0.589,
+		hle: 0.042,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 8B (Reasoning)",
+	},
+	"qwen3-max-thinking-preview": {
+		// AA specific benchmarks
+		codingIndex: 24.5,
+		mathIndex: 82.3,
+
+		// Academic benchmarks
+		mmluPro: 0.824,
+		gpqa: 0.776,
+		hle: 0.12,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Max Thinking (Preview)",
+	},
+	"qwen1.5-chat-110b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.289,
+		hle: undefined,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen1.5 Chat 110B",
+	},
+	"qwen2.5-instruct-32b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.697,
+		gpqa: 0.466,
+		hle: 0.038,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Instruct 32B",
+	},
+	"qwen3-vl-30b-a3b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 14.3,
+		mathIndex: 72.3,
+
+		// Academic benchmarks
+		mmluPro: 0.764,
+		gpqa: 0.695,
+		hle: 0.064,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 30B A3B Instruct",
+	},
+	"qwen3.5-27b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 33.4,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.842,
+		hle: 0.132,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3.5 27B (Non-reasoning)",
+	},
+	"qwen3-14b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 13.1,
+		mathIndex: 55.7,
+
+		// Academic benchmarks
+		mmluPro: 0.774,
+		gpqa: 0.604,
+		hle: 0.043,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 14B (Reasoning)",
+	},
+	"qwen3-30b-a3b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 11,
+		mathIndex: 72.3,
+
+		// Academic benchmarks
+		mmluPro: 0.777,
+		gpqa: 0.616,
+		hle: 0.066,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 30B A3B (Reasoning)",
+	},
+	"qwen3-0.6b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 1.4,
+		mathIndex: 10.3,
+
+		// Academic benchmarks
+		mmluPro: 0.231,
+		gpqa: 0.231,
+		hle: 0.052,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 0.6B (Non-reasoning)",
+	},
+	"qwen3-vl-8b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 7.3,
+		mathIndex: 27.3,
+
+		// Academic benchmarks
+		mmluPro: 0.686,
+		gpqa: 0.427,
+		hle: 0.029,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 8B Instruct",
+	},
+	"qwen3-4b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: 22.3,
+
+		// Academic benchmarks
+		mmluPro: 0.696,
+		gpqa: 0.522,
+		hle: 0.051,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 4B (Reasoning)",
+	},
+	"qwen3-1.7b-non-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 2.3,
+		mathIndex: 7.3,
+
+		// Academic benchmarks
+		mmluPro: 0.411,
+		gpqa: 0.283,
+		hle: 0.052,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 1.7B (Non-reasoning)",
+	},
+	"qwen3-vl-32b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 14.5,
+		mathIndex: 84.7,
+
+		// Academic benchmarks
+		mmluPro: 0.818,
+		gpqa: 0.733,
+		hle: 0.096,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 32B (Reasoning)",
+	},
+	"qwen3-30b-a3b-2507-instruct": {
+		// AA specific benchmarks
+		codingIndex: 14.2,
+		mathIndex: 66.3,
+
+		// Academic benchmarks
+		mmluPro: 0.777,
+		gpqa: 0.659,
+		hle: 0.068,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 30B A3B 2507 Instruct",
+	},
+	"qwen3-30b-a3b-2507-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 14.6,
+		mathIndex: 56.3,
+
+		// Academic benchmarks
+		mmluPro: 0.805,
+		gpqa: 0.707,
+		hle: 0.098,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 30B A3B 2507 (Reasoning)",
+	},
+	"qwen-chat-72b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: undefined,
+		hle: undefined,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen Chat 72B",
+	},
+	"qwen3-vl-32b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 15.6,
+		mathIndex: 68.3,
+
+		// Academic benchmarks
+		mmluPro: 0.791,
+		gpqa: 0.671,
+		hle: 0.063,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 VL 32B Instruct",
+	},
+	"qwen3-0.6b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 0.9,
+		mathIndex: 18,
+
+		// Academic benchmarks
+		mmluPro: 0.347,
+		gpqa: 0.239,
+		hle: 0.057,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 0.6B (Reasoning)",
+	},
+	"qwen3-1.7b-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 1.4,
+		mathIndex: 38.7,
+
+		// Academic benchmarks
+		mmluPro: 0.57,
+		gpqa: 0.356,
+		hle: 0.048,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 1.7B (Reasoning)",
+	},
+	"qwen3-max-preview": {
+		// AA specific benchmarks
+		codingIndex: 25.5,
+		mathIndex: 75,
+
+		// Academic benchmarks
+		mmluPro: 0.838,
+		gpqa: 0.764,
+		hle: 0.093,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Max (Preview)",
+	},
+	"qwen3.6-max-preview": {
+		// AA specific benchmarks
+		codingIndex: 44.9,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.888,
+		hle: 0.289,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3.6 Max Preview",
+	},
+	"qwen3-max": {
+		// AA specific benchmarks
+		codingIndex: 26.4,
+		mathIndex: 80.7,
+
+		// Academic benchmarks
+		mmluPro: 0.841,
+		gpqa: 0.764,
+		hle: 0.111,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Max",
+	},
+	"qwen3-max-thinking": {
+		// AA specific benchmarks
+		codingIndex: 30.5,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: undefined,
+		gpqa: 0.861,
+		hle: 0.262,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 Max Thinking",
+	},
+	"qwen3-4b-2507-reasoning": {
+		// AA specific benchmarks
+		codingIndex: 9.5,
+		mathIndex: 82.7,
+
+		// Academic benchmarks
+		mmluPro: 0.743,
+		gpqa: 0.667,
+		hle: 0.059,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 4B 2507 (Reasoning)",
+	},
+	"qwen3-4b-2507-instruct": {
+		// AA specific benchmarks
+		codingIndex: 9,
+		mathIndex: 52.3,
+
+		// Academic benchmarks
+		mmluPro: 0.672,
+		gpqa: 0.517,
+		hle: 0.047,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen3 4B 2507 Instruct",
+	},
+	"qwen2.5-coder-instruct-7b": {
+		// AA specific benchmarks
+		codingIndex: undefined,
+		mathIndex: undefined,
+
+		// Academic benchmarks
+		mmluPro: 0.473,
+		gpqa: 0.339,
+		hle: 0.048,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Qwen2.5 Coder Instruct 7B ",
+	},
+	"ling-1t": {
+		// AA specific benchmarks
+		codingIndex: 18.8,
+		mathIndex: 71.3,
+
+		// Academic benchmarks
+		mmluPro: 0.822,
+		gpqa: 0.719,
+		hle: 0.072,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Ling-1T",
+	},
+	"ling-flash-2.0": {
+		// AA specific benchmarks
+		codingIndex: 16.7,
+		mathIndex: 65.3,
+
+		// Academic benchmarks
+		mmluPro: 0.777,
+		gpqa: 0.657,
+		hle: 0.063,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Ling-flash-2.0",
+	},
+	"ring-1t": {
+		// AA specific benchmarks
+		codingIndex: 16.8,
+		mathIndex: 89.3,
+
+		// Academic benchmarks
+		mmluPro: 0.806,
+		gpqa: 0.774,
+		hle: 0.102,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Ring-1T",
+	},
+	"seed-oss-36b-instruct": {
+		// AA specific benchmarks
+		codingIndex: 16.7,
+		mathIndex: 84.7,
+
+		// Academic benchmarks
+		mmluPro: 0.815,
+		gpqa: 0.726,
+		hle: 0.091,
+
+		// Capabilities
+		contextWindow: 8192,
+		supportsReasoning: false,
+		supportsVision: false,
+
+		// Metadata
+		lastUpdated: "2026-06-01",
+		originalModel: "Seed-OSS-36B-Instruct",
+	},
+};