Hy4 preview: benchmark results
Tencent's open-weight Hy4 preview (August 2026), a 770-billion-parameter Mixture-of-Experts model from the Tencent Hy team, activating 49B parameters per token with a built-in speculative-decoding module. Provider: Tencent. Released 2026-08-28. Access: Open.
Aggregate Rating 1838 ± 9, rank #38 of 1656 rated models, from 249 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM Stats (CritPT) | 16.9 | Score (%) | 100 |
| LLM Stats (MathArena Apex) | 74.2 | Score (%) | 100 |
| LLM Stats (SuperChem) | 66.4 | Score (%) | 100 |
| LLM Stats (WideSearch) | 83.9 | Score (%) | 100 |
| Nejumi 4 - ARC-AGI-1 | 100 | Accuracy (%) | 100 |
| Nejumi 4 - jaster (0-shot) - JSQuAD | 83.3 | Character F1 (x100) | 100 |
| Nejumi 4 - jaster (2-shot) - MMLU-ProX (Japanese) | 92 | Exact match (%) | 100 |
| WideResearch (BenchLM) | 83.9 | Task Score (%) | 100 |
| ReLE - Language and Instruction Following | 78.8 | Accuracy (%) | 99.5 |
| Nejumi 4 - jaster (2-shot) - ALT English-to-Japanese | 92.72 | COMET wmt22 (x100) | 99.3 |
| Nejumi 4 - GLP - Information Retrieval | 82.3 | Score (%) | 97.8 |
| ReLE - Education - Middle School Subjects | 58.8 | Accuracy (%) | 97.1 |
Interactive version: theaggregate.ai/model?slug=hy4-preview · How It Works · Data refreshed daily, snapshot 2026-10-20.