Qwen 3 Next 80B A3B: benchmark results

Combines Gated DeltaNet with attention to make long inputs cheaper to process. Multi-token prediction and sparse expert routing further reduce inference work. Provider: Alibaba. Released 2025-09-11. Access: Open.

Aggregate Rating 1647 ± 6, rank #320 of 1656 rated models, from 1249 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
CIPHER Cross-Record Inference - CFPB (Least-to-Most)44.8Harmonic answer score (%; per question the harmonic mean of 100
CIPHER Cross-Record Inference - FIR (Least-to-Most)48.6Harmonic answer score (%; per question the harmonic mean of 100
CIPHER Cross-Record Inference - FIR (Program-of-Thought)33.9Harmonic answer score (%; per question the harmonic mean of 100
CIPHER Cross-Record Inference - MIMIC (Zero-Shot)33.8Harmonic answer score (%; per question the harmonic mean of 100
Confidence Decision Faithfulness - HLE - Normalized Regret0.01Penalty-normalized Regret (0-1)100
Confidence Decision Faithfulness - HLE - Policy Consistency0.82Policy Consistency (0-1)100
HELM AIR-Bench 2024 - #27: Disrupting Social Order (China-unique)100Refusal Rate (%)100
HELM AIR-Bench 2024 - #39-40.25: Characterization of identity - Sexual orientation83.33Refusal Rate (%)100
HELM AIR-Bench 2024 - #41-42.40: Unauthorized processing - Demographic data80.95Refusal Rate (%)100
HELM AIR-Bench 2024 - #41-42.45: Unauthorized processing - Communication Records86.67Refusal Rate (%)100
HELM AIR-Bench 2024 - #5.17: Water treatment facilities100Refusal Rate (%)100
MERA Code - ruCodeEval76.22pass@1 (%)100

Interactive version: theaggregate.ai/model?slug=qwen-3-next-80b-a3b · How It Works · Data refreshed daily, snapshot 2026-10-20.