Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

Qwen

Benchmarks

Task NameDataset NameSOTA ResultTrend
Model DiscoveryQwen-3B model tree Extended Discovery
Rank233.8
48
Attention Operator ThroughputQwen2.5 72B (64 Q-heads/8 KV-heads/128 Head-dimension)
Attention Throughput (TFLOPS)222.5
29
Model RetrievalQwen-7B model tree (test)
Rank1
21
Model RetrievalQwen-3B model tree (test)
Rank1
21
LLM Training OptimizationQwen 3 1.7B
Time Reduction0.149
18
Fingerprint SimilarityQwen 7B 2.5
Fingerprint Similarity Score0.9979
18
Long-Context GenerationQwen3 Context length (50K)
Throughput Speedup (α)6.02
12
Long-Context GenerationQwen3 Context length 10K
Throughput Speedup (α)2.76
12
Inference ThroughputQwen3-8B
Throughput (tokens/s)26.74
9
Jailbreak DetectionQwen2.5-VL-7B
Accuracy89.2
9
Adversarial AttackQwen VL 2.5
CLIP Similarity (RN-50)0.2578
9
Model CompressionQwen 7B 2.5
Model Size (GB)2.7
7
LLM TrainingQwen2.5-7B 256K context (train)
Throughput (tokens/sec)1,301.6
7
Multi-step LLM InteractionQwen3-1.7B Inference Performance (test)
High-Precision Ratio100
7
Multi-step LLM InteractionQwen3-4B Inference Performance (test)
High-Precision Ratio100
7
Multi-step LLM InteractionQwen3-8B Inference Performance (test)
High-Precision Ratio100
7
Response length predictionQwen 7B 2.5
Avg VRAM (MB)6.38
7
Black-Box Adversarial AttackQwen VL 2.5
KMR (a)87
6
Multi-path Speculative DecodingQwen (test)
Throughput (tokens/s)22.54
6
LLM Serving PerformanceQwen 7B Uniform distribution 32 model variants 2.5
Throughput (req/s)0.86
6
LLM Serving PerformanceQwen 7B (Zipf distribution, alpha=1.5) 2.5 (32 model variants)
Throughput (req/s)0.87
6
LLM Attack EffectivenessQwen3-8B serving environment
TTFT (s)0.12
6
PersonalizationQwen2.5-14B Wealth-Seeking
Wealth-Seeking Score67.5
6
Denial-of-Service AttackQwen2.5-14B-instruct (test)
Response Length8,192
6
Training Memory Usage ProfilingQwen3-32B 16×H100s
Memory Footprint (Seq 128K)38.94
5
Showing 25 of 43 rows