Hunyuan-4B-Pretrain: benchmark results

Provider: Tencent. Access: Open.

Aggregate Rating 1490 ± 37, rank #2455 of 10552 rated models, from 26 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Lara Mini Base Report (base-model protocol) - CRUXEval-I (2-shot)56.3pass@1, direct prediction, official answer parser (%)100
Lara Mini Base Report (base-model protocol) - CRUXEval-O (2-shot)57pass@1, direct prediction, official answer parser (%)100
Lara Mini Base Report (base-model protocol) - MATH (4-shot)74.8Math-Verify equivalence, Minerva prompts, 5,000 problems ove100
Lara Mini Base Report (base-model protocol) - MMLU-Pro (5 (CoT)-shot)55.1Exact match (%), English benchmark98.1
Lara Mini Base Report (base-model protocol) - MGSM (8 (CoT)-shot)63.1Exact match on each language's native CoT answer phrase, ave96.3
Lara Mini Base Report (base-model protocol) - MBPP+ (3-shot)62.4pass@1; must pass MBPP's own and the EvalPlus tests (%)94.4
Lara Mini Base Report (base-model protocol) - MBPP (3-shot)58.8pass@1 (execution) (%)93.5
Lara Mini Base Report (base-model protocol) - PAWS-X, All Languages (0-shot)59.7Accuracy (%) averaged over all languages/variants of the ben93.5
Lara Mini Base Report (base-model protocol) - Belebele (English) (0-shot)87.3Accuracy (%), English benchmark92.6
Lara Mini Base Report (base-model protocol) - MMLU (5-shot)72.5Accuracy (%), English benchmark92.6
Lara Mini Base Report (base-model protocol) - GSM8K (5-shot)78Exact match (strict extraction: first "#### N", else first \90.7
Lara Mini Base Report (base-model protocol) - Belebele, All Languages (0-shot)49Accuracy (%) averaged over all languages/variants of the ben88.9

Interactive version: theaggregate.ai/model?slug=hunyuan-4b-pretrain · How It Works · Data refreshed daily, snapshot 2026-11-16.