Qwen3 8 vs Qwen3.8 27B
Qwen3 8 Verdict
- Qwen3 8 (max) leads on quality: 66.5 vs 59.2.
- Qwen3 8 (max) is stronger in agents & tools, coding, composite, human preference, knowledge, maths, multimodal, reasoning.
- Qwen3.8 27B (xhigh) is stronger in long context.
- Qwen3.8 27B (xhigh) is 2.8× cheaper ($0.938 vs $2.67 per 1M blended).
| Metric | Qwen3 8 (max) | Qwen3.8 27B (xhigh) |
|---|---|---|
| BenchLeader Index | 66.5 | 59.2 |
| Agents & tools score | 68.3 | 67.3 |
| Coding score | 66.5 | 53.2 |
| Composite score | 74.3 | 72.0 |
| Human preference score | 68.2 | – |
| Knowledge score | 63.5 | 56.4 |
| Long context score | 66.6 | 67.5 |
| Maths score | 67.0 | – |
| Multimodal score | 67.4 | 60.7 |
| Reasoning score | 76.3 | 52.5 |
| Blended price $/M | $2.67 | $0.938 |
| Output speed | 39 tok/s | 42 tok/s |
| Time to first answer | 55.1 s | 51.0 s |
| Context window | 1M | 262k |
| SciCode | 52.9% | 44.7% |
| ProofBench | 58.0% | – |
| Epoch Capabilities Index | 156.6 | – |
| LMArena Text | 1481 | – |
| LMArena Hard Prompts | 1503 | – |
| LMArena Coding | 1522 | – |
| LMArena WebDev | 1671 | – |
| LMArena Vision | 1315 | – |
| LMArena Agent | 3.3 | – |
| LiveBench | 78.5% | – |
| LiveBench Reasoning | 88.2% | – |
| LiveBench Coding | 72.9% | – |
| LiveBench Agentic Coding | 64.7% | – |
| LiveBench Mathematics | 91.3% | – |
| LiveBench Data Analysis | 78.4% | – |
| LiveBench Language | 79.7% | – |
| LiveBench Instruction Following | 74.1% | – |
| AA Intelligence Index | 45.4 | 33.9 |
| AA-LCR | 80.3% | 82.0% |
| MMMU-Pro | 82.8% | 76.3% |
| AA-Omniscience | 12.0 | -10.0 |
| GPQA Diamond (AA) | 92.8% | 90.5% |
| Humanity's Last Exam (AA) | 43.1% | 33.9% |
| SciCode (AA) | 53.2% | 46.6% |
| LiveCodeBench | 87.8% | 84.0% |
| MMLU-Pro | 88.6% | 84.3% |
| IOI | 68.9% | 39.1% |
| LegalBench | 83.6% | 82.4% |
| CorpFin | 65.8% | – |
| TaxEval | 75.5% | 70.8% |
| Terminal-Bench 2.1 (Vals) | 67.4% | 58.4% |
| SWE-bench (Vals) | 85.6% | 86.0% |
| GPQA Diamond (Vals) | 93.7% | 88.9% |
| Vals Index | 51.8 | 48.5 |
| CritPt | 20.0% | 5.4% |
| GDPval (AA) | 58.2% | 48.2% |
| τ²-Bench Banking (AA) | 51.3% | 48.0% |
| Code Migration | 24.0% | 14.2% |
| Excel Modeling Benchmark | 60.1% | 59.7% |
| Finance Agent v2 | 50.6% | 48.5% |
| Harvey's Legal Agent Benchmark | 10.4% | 11.3% |
| Legal Research Bench | 47.6% | 36.1% |
| MedCode | 40.7% | 28.7% |
| MedScribe | 85.0% | 83.8% |
| MMMU-Pro (Vals) | 88.0% | 83.9% |
| MortgageTax | 64.0% | 64.9% |
| MysteryMechanism | 23.9% | – |
| ProgramBench | 0.0% | 0.0% |
| Public Benefits Bench | 67.1% | – |
| SAGE | 51.3% | 52.4% |
| SkillsBench | 42.0% | 38.1% |
| Tax Agent Bench | 66.0% | – |
| Terminal-Bench 4.0 (Vals) | 24.8% | 4.0% |
| Terminal-Bench Science | 1.4% | 1.4% |
| Vals Multimodal Index | 65.4% | – |
| Vibe Code Bench 1-100 | 12.8% | – |
| Vibe Code Bench v1.1 | 64.7% | 64.8% |
| LMArena Maths | 1498 | – |
| LMArena Creative Writing | 1472 | – |
| LMArena Instruction Following | 1477 | – |
| LMArena Multi-turn | 1495 | – |
| LMArena Longer Queries | 1492 | – |
| Surface Evolver Bench | – | 45.0% |
Data as of 2026-09-19. Best configuration of each model; every score links to its source on the model pages.
Qwen3 8 vs Qwen3.8 27B: questions
- Is Qwen3 8 better than Qwen3.8 27B?
- Qwen3 8 (max) leads on quality: 66.5 vs 59.2. The BenchLeader Index combines every independent quality benchmark; Qwen3 8 (max) is ahead overall as of 2026-09-19, but check the category scores for your use.
- Is Qwen3 8 better than Qwen3.8 27B for coding?
- Qwen3 8 scores higher in coding (67 vs 53 on the category index, where 50 is average).
- Is Qwen3 8 better than Qwen3.8 27B for agentic tasks?
- Qwen3 8 scores higher in agentic tasks (68 vs 67 on the category index, where 50 is average).
- Which is cheaper, Qwen3 8 or Qwen3.8 27B?
- Qwen3.8 27B is cheaper: $0.938 against $2.67 per million tokens, blended at three input tokens per output token.
- Which is faster, Qwen3 8 or Qwen3.8 27B?
- Qwen3.8 27B streams faster: 42 against 39 output tokens per second.
- Which has the larger context window?
- Qwen3 8 accepts more context: 1M against 262k tokens.