SupermaginSupermagin
Qwen3-Max

Qwen3-Max

qwen· qwenwarning

Qwen3-Max: curated reference profile, Apache-2.0 license

Qwen3-Max: curated reference profile drawn from the model's published reports. Not a Supermagin measurement. Run the official eval in the Studio to earn a verified score.

Qwen+8% net improvement

Curated reference profile · not a Supermagin measurement

Model card & documentation

Qwen3-Max

Curated reference profile · Apache-2.0 license · v3.0 · 1.2s latency

These figures are drawn from the model's published reports and are NOT a Supermagin measurement. Run the official eval in the Studio to earn a verified, re-verifiable score.

Capabilities

  • coding: 90%
  • reasoning: 88%
  • agent: 87%
  • cybersecurity: 85%

Notes

  • Low hallucination rate on audited evals.
  • Reference evals: SWE-Bench Verified, GPQA Diamond, MMLU-Pro, MATH-500, ARC-AGI-2, BigCodeBench, τ-bench.

Pooled score

0%

credibility 100% · consistency 0.96

Score trend across attested runs

24880 total samples

More attested runs are needed to show a trend.

Category scores

reasoning
88%
coding
90%
agent
87%
multimodal
84%
cybersecurity
85%

Progress by year

2023
78%
2024
85%
2025
89%
2026
91%

Success rate

0%

Failure rate

0%

Hallucination

0%

Avg latency

1200ms

per task

Avg cost

$0.0030

per task

Calibration gap

+0.12

positive = overconfident

Top failure signals

hallucination ×697logic_flaw ×1234instruction_failure ×926

Pooled brain diagnostics

Output Entropy
healthy
Top-Token Confidence
healthy
Participation Ratio
healthy
Condition Number
healthy
Gradient Chain Health
healthy

Failure rate by task type

code
15%8708
qa
10%7464
reasoning
12%4976
classification
11%3732

Curated reference profile, descriptive data, not a Supermagin measurement.