Qwen
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
Qwen3-4B
100Mean Certifiable Acceptance Length
72
Qwen3-8B Single-layer forward+backward setup
36.6Time (ms)
57
Qwen-1.5B-Instruct 2.5 (test)
93Score S(γ)
56
Qwen-3B model tree Extended Discovery
233.8Rank
48
Qwen2-VL
0ASR
36
Qwen2-VL
0.05Toxicity Score
36
Qwen3-8B 2k prompts Decode-heavy workload
46,782Throughput (tok/s)
30
Qwen3 Query Projection Module NVIDIA A40
80.63Throughput (k tokens/sec)
30
Qwen2.5 72B (64 Q-heads/8 KV-heads/128 Head-dimension)
222.5Attention Throughput (TFLOPS)
29
Qwen3-8B 2k prompts Balanced workload
61,106Throughput (tok/s)
28
Qwen 7B 2.5
1,847Training Throughput (tokens/s)
28
Qwen3-8B 2k prompts Prefill-heavy workload
106,324Throughput (tok/s)
26
Qwen3 1.7B
58.65Average Performance
24
Qwen 7B-Instruct 2.5
64.6L(F)
24
Qwen 3B Instruct 2.5
56.9L(F)
24
Qwen2.5-1.5B-Instruct
31.4L(F)
24
Qwen3-1.7B target τpool=1.509 (test)
53.9Tau Multiplier (×τ)
22
Qwen-7B model tree (test)
1Rank
21
Qwen-3B model tree (test)
1Rank
21
Qwen2.5-7B
0.02Normalized Rate (NR)
20
Qwen Target Transferability set 27B
38ASR
19
QWEN3-like (train)
2.398Loss
19
Qwen 1.7B 3
21.69Input Success Rate (%)
18
Qwen3-80B Large Target
99Similarity Score
18
Qwen3-30B Large Target
99Similarity Score
18