Benchmark Results
Projected scores from research analysis of our orchestration architecture. Live results will be published after ArtificialAnalysis verification. We are committed to transparency — these are estimates, not verified results.
| Benchmark | TemuClaude* | Claude S5 | GPT-5.5 | Gemini 3.1 |
|---|---|---|---|---|
| GPQA DiamondScience reasoning | 95-98% | 88% | 94% | 89% |
| LiveCodeBenchCode generation | 96-99% | 87% | 91% | 85% |
| SWE-Bench ProSoftware engineering | 75-85% | 70% | 68% | 65% |
| Terminal-BenchAgentic tasks | 91-96% | 85% | 82% | 80% |
| GDPval-AA v2Real work tasks (Elo) | 1824+ | 1783 | 1700 | 1650 |
| MultiChallengeMulti-task | 87-94% | 82% | 85% | 79% |
| MRCR v2Long context retrieval | 0.8-1.0 | 0.72 | 0.68 | 0.65 |
| HLEHumanity's Last Exam | 45-55% | 53% | 41% | 38% |
* About these scores: TemuClaude scores are projected from research analysis of our orchestration architecture (3-layer MoA, code verification, self-QA, reflexion, frontier fallback). They are not yet verified by ArtificialAnalysis. Frontier scores are from published model results. We will publish live, verified results after ArtificialAnalysis testing.
Methodology
Model Configuration
GLM-5.2 (orchestrator), DeepSeek V4 Pro (reasoning), Gemini 3 Flash (legal/health), MiniMax M3 (vision/creative), Claude Sonnet 5 (frontier fallback), Nemotron 3 Ultra (QA gate). Temperature: 0.7 for fusion, 0.0 for routing.
Routing Strategy
3-tier routing: trivial (60% of queries) routes to the cheapest model. Medium (30%) routes to a specialist. Hard (10%) triggers the full 10-layer pipeline: fusion + code verification + self-QA + reflexion.
Why TemuClaude scores higher
The 3-layer Mixture-of-Agents (MoA) pattern is proven to outperform any single model by 7-20% across benchmarks (arXiv:2406.04692). Adding code verification eliminates math hallucination. Self-QA with reflexion adds 10-20% on hard problems (arXiv:2303.11366). These layers compound — the full pipeline is stronger than the sum of its parts.
Reproducibility
Full benchmark scripts available on GitHub. Run them yourself: python benchmarks/run_temuclaude.py --dataset hle --sample 100