Muse Spark 1.2 vs Qwen3 8
Muse Spark 1.2 
Qwen3 8 Verdict
- Qwen3 8 (max) leads on quality: 63.9 vs 62.6.
- Muse Spark 1.2 is stronger in composite, knowledge, long context, maths, reasoning.
- Qwen3 8 (max) is stronger in coding, agents & tools, human preference, multimodal.
- Muse Spark 1.2 is 1.3× cheaper ($2.00 vs $2.67 per 1M blended).
- Muse Spark 1.2 streams 5.6× faster (219 vs 39 tokens per second).
| Metric | Muse Spark 1.2 | Qwen3 8 (max) |
|---|---|---|
| BenchLeader Index | 62.6 | 63.9 |
| Coding score | 66.5 | 71.7 |
| Composite score | 79.1 | 70.8 |
| Knowledge score | 76.9 | 62.3 |
| Long context score | 66.0 | 65.7 |
| Maths score | 54.3 | – |
| Reasoning score | 70.2 | 67.7 |
| Agents & tools score | – | 56.1 |
| Human preference score | – | 67.9 |
| Multimodal score | – | 67.2 |
| Blended price $/M | $2.00 | $2.67 |
| Output speed | 219 tok/s | 39 tok/s |
| Time to first answer | 23.7 s | 53.2 s |
| Context window | 1.0M | 1M |
| SimpleBench | 74.5% | – |
| ProofBench | 43.0% | – |
| Epoch Capabilities Index | 155.5 | 156.6 |
| LMArena Text | – | 1480 |
| LMArena Hard Prompts | – | 1502 |
| LMArena Coding | – | 1520 |
| LMArena WebDev | – | 1685 |
| LMArena Vision | – | 1313 |
| LMArena Agent | – | 3.9 |
| 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% |
| AA Intelligence Index | 39.8 | 40.3 |
| AA-LCR | 79.0% | 78.3% |
| MMMU-Pro | – | 82.3% |
| AA-Omniscience | 27.2 | 3.4 |
| GPQA Diamond (AA) | 90.4% | 92.7% |
| Humanity's Last Exam (AA) | 45.5% | 43.0% |
| SciCode (AA) | 57.4% | 53.2% |
| LiveCodeBench | – | 87.8% |
| MMLU-Pro | – | 88.6% |
| IOI | 49.5% | 73.0% |
| LegalBench | – | 83.6% |
| CorpFin | – | 65.8% |
| TaxEval | – | 75.5% |
| Terminal-Bench 2.1 (Vals) | – | 67.4% |
| SWE-bench (Vals) | – | 85.6% |
| GPQA Diamond (Vals) | – | 93.7% |
| Vals Index | – | 51.8 |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.