Alibaba · Qwen3.8
Qwen3.8-Flash-Next
Open preview of the architecture planned for Qwen4 (125B MoE with 6B active, plus 51B n-gram embedding and 4B MTP parameters, about 180B total), and the base of the hosted Qwen3.8-Flash.
- AA Intelligence
- 40#21 of 142
- Price
- Self-host
- Output speed
- 54tok/s
- Context window
- 262K tokens
Composite indexes
- Artificial Analysis Intelligence Index#21 of 14240
Leader: Claude Opus 5.5 at 58
Coding
- Terminal-Bench 4.0#19 of 8825.3%
Leader: Claude Mythos 5.1 at 60.9%
- Terminal-Bench 2.x#12 of 10386.1%
Leader: Claude Fable 5.1 at 91.4%
- SciCode#37 of 9450.6%
Leader: Claude Opus 5.5 at 66.9%
- SWE-bench Pro#9 of 3162.5%
Leader: Claude Opus 5.5 at 89.9%
- LiveCodeBench#2 of 5691.9%
Leader: Fugu at 92.9%
Agents and tool use
- GDPval-AA#10 of 1031612
Leader: Claude Opus 5.5 at 1846
Reasoning and knowledge
- GPQA Diamond#22 of 14292.3%91.7%
Leader: GPT-6 Astra at 96.1%
- Humanity’s Last Exam#42 of 14738%35.9%
Leader: Claude Opus 5.5 at 61.4%
- CritPt#36 of 13911.1%
Leader: GPT-5.6 Sol at 32.3%
Vision and long context
- AA Long Context Reasoning#34 of 13979.7%
Leader: Kimi K3 at 88.7%
- MMMU-Pro#18 of 7279.8%
Leader: Claude Opus 5.5 at 87.7%
Human preference
- LMArena WebDev#9 of 541636
Leader: Claude Opus 5.5 at 1818
Other published results
| Result | Score | Reported by | Source |
|---|---|---|---|
| DeepSWE 1.1 | 58.7% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |
| SWE-bench Multilingual | 81% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |
| IFBench | 81.3% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |
| Toolathlon Verified · Pass@1 | 73.5% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |
| OSWorld 2.0 (binary) · partial-credit score 52.3 | 19.4% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |
| AndroidWorld | 84.5% | vendor | Qwen3.8-Flash-Next model card (Hugging Face) ↗ |