Qwen-SEA-LION-v5-397B-A17B-Plus: benchmark results

AI Singapore's SEA-LION v5 flagship (Nov 18, 2026, MIT), converting Qwen3.5-397B-A17B to Qwen4Exp with ~164B host-offloaded n-gram embedding tables (569.45B safetensors total, 17B active). Provider: AI Singapore. Released 2026-11-18. Access: Open.

Aggregate Rating 1751 ± 36, rank #178 of 10524 rated models, from 132 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
SEA-HELM77.1Mean Score (%)100
SEA-HELM (Burmese)70.58Normalized Score100
SEA-HELM (Burmese) - Belebele QA88.01Normalized Score100
SEA-HELM (Burmese) - Causal Reasoning84.77Normalized Score100
SEA-HELM (Burmese) - Global MMLU Lite81.72Normalized Score100
SEA-HELM (Burmese) - NLR70.54Normalized Score100
SEA-HELM (Burmese) - NLU68.09Normalized Score100
SEA-HELM (Filipino) - Global MMLU Lite87.48Normalized Score100
SEA-HELM (Filipino) - Sentiment Analysis72.17Normalized Score100
SEA-HELM (Indonesian)80.29Normalized Score100
SEA-HELM (Malay) - Safety47.68Normalized Score100
SEA-HELM (Malay) - Sentiment Analysis53.52Normalized Score100

Interactive version: theaggregate.ai/model?slug=qwen-sea-lion-v5-397b-a17b-plus · How It Works · Data refreshed daily, snapshot 2026-11-22.