2026-08-02
Qwen3.8-Max
by Qwen
Expected Performance
56.1%
Expected Rank
#12
Expected Cost / Problem
$3.07
Competition performance
| Competition | Accuracy | Rank | Cost | Output Tokens |
|---|---|---|---|---|
|
06/2026
ArXivLean
|
45.65% ± 14.39% | 3/14 | $10.27 | 437952 |
|
Overall
BrokenArXiv
|
48.71% ± 5.93% | 8/12 | $1.52 | 142278 |
|
04/2026
BrokenArXiv
|
45.49% ± 8.84% | 6/19 | $0.35 | 57731 |
|
05/2026
BrokenArXiv
|
32.50% ± 9.18% | 12/23 | $0.39 | 65360 |
|
06/2026
BrokenArXiv
|
43.98% ± 9.36% | 15/28 | $0.38 | 63195 |
|
08/2026
BrokenArXiv
|
69.64% ± 12.04% | 8/13 | $3.64 | 298279 |
|
Overall
ArXivMath
|
67.59% ± 5.59% | 6/12 | $2.49 | 140190 |
|
04/2026
ArXivMath
|
59.17% ± 8.79% | 9/23 | $0.30 | 49511 |
|
05/2026
ArXivMath
|
69.17% ± 8.26% | 10/26 | $0.35 | 58574 |
|
06/2026
ArXivMath
|
75.69% ± 7.01% | 12/29 | $0.39 | 64770 |
|
08/2026
ArXivMath
|
57.89% ± 12.82% | 8/13 | $5.76 | 297226 |
Accuracy
45.65%
Overall BrokenArXiv
Accuracy
48.71%
04/2026 BrokenArXiv
Accuracy
45.49%
05/2026 BrokenArXiv
Accuracy
32.50%
06/2026 BrokenArXiv
Accuracy
43.98%
08/2026 BrokenArXiv
Accuracy
69.64%
Overall ArXivMath
Accuracy
67.59%
04/2026 ArXivMath
Accuracy
59.17%
05/2026 ArXivMath
Accuracy
69.17%
06/2026 ArXivMath
Accuracy
75.69%
08/2026 ArXivMath
Accuracy
57.89%
Sampling parameters
- Model
- qwen/qwen3.8-max
- API
- openrouter
- Display Name
- Qwen3.8-Max
- Release Date
- 2026-08-02
- Open Source
- Yes
- Creator
- Qwen
- Parameters (B)
- 2400
- Active Parameters (B)
- 95
- Max Tokens
- 131072
- Temperature
- 1.0
- Top-p
- 0.95
- Read cost ($ per 1M)
- 2
- Write cost ($ per 1M)
- 6
- Concurrent Requests
- 64
Additional parameters
{
"cache_read_cost": 0.25,
"extra_body": {
"provider": {
"allow_fallbacks": false,
"order": [
"alibaba"
]
},
"top_k": 20
},
"harness": "qwen",
"harness_config": {
"auth": "api",
"container_executable": "qwen",
"environment": {
"QWEN_STREAM_MAX_LIFETIME_MS": "0"
}
},
"harness_version": "0.24.0",
"huggingface_id": "Qwen/Qwen3.8-2.4T-A95B",
"reasoning_effort": "max"
}
Most surprising traces (Item Response Theory)
Computed once using a Rasch-style logistic fit; excludes Project Euler where traces are hidden.
Surprising failures
Click a trace button above to load it.
Surprising successes
Click a trace button above to load it.