Qwen3.8 2.4T A95B: benchmark results
Alibaba's open-weight Qwen3.8-2.4T-A95B (August 2026), the post-trained base of Qwen3.8-Max and the first Qwen-Max-class model released open, a 2.4-trillion-parameter MoE with 95 billion active parameters. Provider: Alibaba. Released 2026-08-12. Access: Open.
Aggregate Rating 1817 ± 11, rank #55 of 1656 rated models, from 259 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA Omniscience - Software Engineering (SWE) - Rust | 82 | Accuracy (%) | 99.8 |
| Nejumi 4 - ALT Average | 91.71 | Score (%) | 99.3 |
| Nejumi 4 - JTruthfulQA | 77.62 | Score (%) | 99.3 |
| Nejumi 4 - jaster (2-shot) - NIILC | 62.84 | Character F1 (x100) | 99.3 |
| Nejumi 4 - jaster (0-shot) - JaMP | 83 | Exact match (%) | 98.6 |
| Nejumi 4 - ALT - Bias | 97.44 | Score (%) | 98.5 |
| Nejumi 4 - Total Score | 85.98 | Score (%) | 98.5 |
| Nejumi 4 - jaster (2-shot) - JCoLA (out-of-domain) | 90 | Exact match (%) | 98.3 |
| Tau3 Banking | 49.07 | Success Rate (%) | 98.1 |
| Humanity's Last Exam (Self-Reported, No Tools) | 43.6 | Accuracy (%) | 98 |
| Nejumi 4 - GLP - Reasoning | 94.11 | Score (%) | 97.8 |
| AA GPQA Diamond | 93.54 | Accuracy (%) | 97.4 |
Interactive version: theaggregate.ai/model?slug=qwen3-8-2-4t-a95b · How It Works · Data refreshed daily, snapshot 2026-10-20.