Hy3 (High): benchmark results
Tencent Hy3 uses 80 layers, with one dense feed-forward layer followed by 79 MoE layers routing to eight of 192 experts plus one shared expert. The publisher lists a 295B backbone, 21B active parameters and a separate 3.8B multi-token-prediction module; complete-checkpoint tensor totals use a different scope. Provider: Tencent. Released 2026-07-06. Access: Open.
Aggregate Rating 1753 ± 14, rank #228 of 2169 rated models, from 95 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ObviousBench | 100 | Answer pass³ (%) | 97.7 |
| Nejumi 4 - MT-Bench (Japanese) - Coding | 100 | Judge rating (1-10, x10) | 91.3 |
| Nejumi 4 - MT-Bench (Japanese) - Humanities | 100 | Judge rating (1-10, x10) | 90.9 |
| Hy-MultiTurn - Precise Execution | 85.5 | Importance-weighted score (%) on the 35 precise execution di | 90.5 |
| Nejumi 4 - JHumanEval | 47.82 | Sandbox pass rate (%) | 90.2 |
| Nejumi 4 - M-IFEval (Japanese) | 89.38 | Score (%) | 88.1 |
| Nejumi 4 - jaster (2-shot) - MGSM | 93 | Exact match (%) | 86.4 |
| Nejumi 4 - BFCL - Multi-Turn | 55.56 | Accuracy (%) | 86 |
| Nejumi 4 - jaster (0-shot) - KUCI | 87 | Exact match (%) | 86 |
| Nejumi 4 - jaster (2-shot) - MMLU-ProX (Japanese) | 89 | Exact match (%) | 85.7 |
| SuperCLUE General (July 2026) - Science Reasoning | 75.44 | Score | 85.3 |
| Nejumi 4 - jaster (2-shot) - JaNLI | 100 | Exact match (%) | 85 |
Interactive version: theaggregate.ai/model?slug=hy3-high · How It Works · Data refreshed daily, snapshot 2026-10-23.