Qwen 3 Coder 480B A35B Instruct: benchmark results
Built for repository-scale coding agents that inspect code and call tools. Uses a dedicated function-call format and answers without separate thinking blocks. Provider: Alibaba. Released 2025-07-22. Access: Open.
Aggregate Rating 1648 ± 8, rank #315 of 1656 rated models, from 261 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ClassEval-Pro | 45.6 | Class-level full-success pass@1 (%): the generated class pas | 100 |
| ClassEval-Pro - Method-Level | 84.2 | Method-level full-success pass@1 (%): share of the class's m | 100 |
| NGDBench - NGD-Fin Aggregation sMAPE | 0.33 | Symmetric mean absolute percentage error (0 to 2) between th | 100 |
| NGDBench - NGD-Fin Management sMAPE | 0.61 | Symmetric mean absolute percentage error (0 to 2) of the num | 100 |
| NGDBench - NGD-Fin Non-Aggregation | 45.6 | F1 (x100) between the result set of the model's Cypher query | 100 |
| NGDBench - NGD-Prime Management sMAPE | 0.31 | Symmetric mean absolute percentage error (0 to 2) of the num | 100 |
| HeaRTS - Cross-modal Translation | 66 | Mean task score (x100) over HeaRTS's cross-modal translation | 93.3 |
| MERA Code - RealCode | 41.15 | pass@1 (%) | 86.2 |
| GeoCodeBench - Geometric Logic Routing | 33.2 | Unit-test pass rate (%) on the 11 geometric logic routing pr | 85.7 |
| TuRTLe Module Completion (NotSoTiny) | 23.43 | Functionality EQV (imported) | 85.7 |
| AA-Omniscience Hallucination Rate | 44.93 | Hallucination Rate (%) | 82.9 |
| AA Omniscience - Software Engineering (SWE) - Dart | 28 | Accuracy (%) | 82.1 |
Interactive version: theaggregate.ai/model?slug=qwen-3-coder-480b-a35b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.