Modell-Vergleich
Welche KI ist am besten für Strix Halo? Vergleich von Cloud-APIs und lokalen Modellen — mit Daten von Artificial Analysis, lmspeed.net und Messungen auf Strix Halo (local-llm-benchmarks.dev).
Cloud-API Intelligence Leaderboard
Daten: 58 Modelle · Stand: 31. Oktober 2026 · Intelligence Index v5.0 · Quellen: Artificial Analysis, lmspeed.net, ollamatps.com
| # | Modell | Anbieter | Intelligence ↓ | Speed (tok/s) | Latenz | Kosten | Kontext |
|---|---|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon | 59Sehr stark | 85 | 118.58s | $7.76 | 1M | |
| 2 | Claude Opus 5.5 | Anthropic | 58Sehr stark | 97 | 729.62s | $10.04 | 1M |
| 3 | GPT-6 Astra | OpenAI | 57Sehr stark | 78 | 270.03s | $6.24 | 1M |
| 4 | GPT-6.1 Sol | OpenAI | 55Sehr stark | 79 | 315.06s | $1.27 | 1M |
| 5 | Claude Sonnet 5.5 | Anthropic | 54Sehr stark | 139 | 463.03s | $10.02 | 1M |
| 6 | Claude Opus 5.5 (high) | Anthropic | 52Sehr stark | 80 | 34.52s | $3.46 | 1M |
| 7 | GPT-6.1 Terra | OpenAI | 51Sehr stark | 97 | 274.62s | $0.96 | 1M |
| 8 | Gemini 4 Argon | 50Sehr stark | 83 | 49.13s | $7.68 | 1M | |
| 9 | GPT-6.1 Sol (high) | OpenAI | 50Sehr stark | 60 | 44.36s | $0.59 | 1M |
| 10 | Muse Wonder Full | Meta | 50Sehr stark | — | — | — | 1M |
| 11 | Claude Fable 5.1 | Anthropic | 49Stark | 69 | 238.66s | $12.99 | 1M |
| 12 | Kimi K3.1 Pro | Moonshot AI | 46Stark | 31 | 68.04s | $3.68 | 1M |
| 13 | Muse Wonder | Meta | 43Stark | 143 | 44.07s | $6.95 | 1.05M |
| 14 | Claude Haiku 5.5 | Anthropic | 36Solide | 235.3 | 341.05s | $0.41 | 1M |
| 15 | Grok 4.7 | SpaceXAI | 34Solide | 87 | 42.06s | $7.51 | 500k |
| 16 | DeepSeek V4.1 Flash 1017 | DeepSeek | 32Solide | 158.6 | 13.52s | — | 1M |
| 17 | GLM-5.3 | Z.AI | 32Solide | 76 | — | — | 1M |
| 18 | MiMo-V2.6-Pro | Xiaomi | 30Solide | 49 | — | — | 1M |
| 19 | Ling 3.1 Flash 1019 | InclusionAI | 29Solide | — | — | — | 1.05M |
| 20 | MiniMax-M3.1-Flash | MiniMax | 29Solide | 119 | — | — | 1M |
| 21 | Step 5 Preview | StepFun | 29Solide | 144.9 | 14.49s | — | 1M |
| 22 | Mistral Large 4 | Mistral | 26Solide | 161 | — | — | 1M |
| 23 | Muse Spark Open | Meta | 25Solide | 139 | — | — | 1.05M |
| 24 | Qwen3.8 27B | Alibaba | 20Einsteiger | 38 | — | — | 256k |
| 25 | K2 Horizon 375B A23B | Institute of Foundation Models | 18Einsteiger | 154 | — | — | 524k |
| 26 | Beam 1.0 | Reflection | 18Einsteiger | 299 | — | — | 1.05M |
| 27 | Nemotron 3.5 Super VL | NVIDIA | 14Basic | 430.2 | 5.2s | — | 1M |
| 28 | Gemini 3.5 Flash-Lite | —— | 392.4 | 4.77s | — | — | |
| 29 | Nemotron 3.5 Lightning | NVIDIA | —— | 277.9 | 7.63s | — | — |
| 30 | GPT-OSS 120B | OpenAI | —— | 248.3 | 8.54s | — | — |
| 31 | Nemotron 3 Nano Omni | NVIDIA | —— | 232.8 | 9.09s | — | — |
| 32 | Qwen3.5 Omni Flash | Alibaba | —— | 225.9 | 0.9s | — | — |
| 33 | Ling 3.1 Flash | InclusionAI | —— | 216.9 | 10.47s | — | — |
| 34 | Inkling Small | Thinking Machines | —— | 187.2 | 12.1s | — | — |
| 35 | Muse Spark 1.3 | Meta | —— | 176.5 | 29.73s | — | — |
| 36 | Inkling | Thinking Machines | —— | 152.9 | 15.08s | — | — |
| 37 | Mistral Medium 3.5 | Mistral | —— | 151.5 | 13.86s | — | — |
| 38 | Quasar 438B | Multiverse Computing | —— | 149.8 | 13.9s | — | — |
| 39 | Command A Plus | Cohere | —— | 147.8 | 13.74s | — | — |
| 40 | GPT-OSS 20B | OpenAI | —— | 147.7 | 14.02s | — | — |
| 41 | Qwen3.5-122B-A10B | Alibaba | —— | 143.8 | 0.96s | — | — |
| 42 | Nemotron 3 Ultra | NVIDIA | —— | 141.7 | 16.87s | — | — |
| 43 | Nemotron 3 Nano 30B | Ollama Cloud Pro | —— | 217.8 | — | — | — |
| 44 | Gemma4 31B | Ollama Cloud Pro | —— | 216.4 | — | — | — |
| 45 | DeepSeek V4.1 Flash | Ollama Cloud Pro | —— | 202 | — | — | — |
| 46 | GLM 5.2 | Ollama Cloud Pro | —— | 184.4 | — | — | — |
| 47 | Nemotron 3 Super | Ollama Cloud Pro | —— | 161.6 | — | — | — |
| 48 | GLM 5.3 Flash | Ollama Cloud Pro | —— | 158.6 | — | — | — |
| 49 | DeepSeek V4 Pro 0813 | Ollama Cloud Pro | —— | 145 | — | — | — |
| 50 | Kimi K2.7 Code | Ollama Cloud Pro | —— | 115.8 | — | — | — |
| 51 | GLM 5.3 | Ollama Cloud Pro | —— | 112 | — | — | — |
| 52 | MiniMax M3 | Ollama Cloud Pro | —— | 104.3 | — | — | — |
| 53 | Kimi K3 | Ollama Cloud Pro | —— | 101.9 | — | — | — |
| 54 | Beam | Ollama Cloud Pro | —— | 96.3 | — | — | — |
| 55 | MiniMax M3.1 Flash | Ollama Cloud Pro | —— | 85.2 | — | — | — |
| 56 | Mistral Large 3 675B | Ollama Cloud Pro | —— | 70 | — | — | — |
| 57 | Kimi K2.6 | Ollama Cloud Pro | —— | 69.3 | — | — | — |
| 58 | MiniMax M2.7 | Ollama Cloud Pro | —— | 39.9 | — | — | — |
Wichtiger Unterschied
Die Spitzenmodelle liefern in der Cloud etwa 60–140 tok/s, also kaum mehr als Qwen3.8-Flash-Next lokal mit Halogen (rund 60 tok/s, siehe „Lokal gemessen“). Der Unterschied liegt in der Intelligenz, nicht im Tempo. Sehr hohe Werte (200–430 tok/s) erreichen nur kleine oder spezialisierte Modelle. Der Intelligence-Score von Artificial Analysis hängt nicht von der Hardware ab und eignet sich daher für den Vergleich.