llmfind llms should run on premises llmfind llms should run on premisestry it — click a dot: the model opens on arena.ai (or OpenRouter chat when arena does not list it) and this prompt lands in your clipboard, paste it there:
Explain in three sentences why the sky is blue, then give me one surprising fact about it.
| model | intelligence | $ / 1M | tok/s | runs on | since | ||
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 anthropic | ● | 59.1 | 9.33 | 95 | 2026-09 | try | |
| GPT-6 Astra openai | ● | 56.9 | 23.3 | 49 | 2026-09 | try | |
| GPT-6.1 Sol openai | ● | 54.6 | 4.67 | 67 | 2026-09 | try | |
| Claude Sonnet 5.5 anthropic | ● | 54.5 | 4.67 | 97 | 2026-09 | try | |
| GPT-6.1 Terra openai | ● | 51.3 | 2.33 | 78 | 2026-10 | try | |
| Gemini 4 Argon google | ● | 50.4 | 4.67 | 91 | 2026-10 | try | |
| Claude Fable 5.1 anthropic | ● | 50 | 23.3 | 53 | 2026-09 | try | |
| Kimi K3.1 Pro moonshotai | ● | 46.5 | 4.67 | 27 | api price | 2026-10 | try |
| GPT-6 Sol openai | ● | 44.2 | 4.67 | 129 | 2026-09 | try | |
| Claude Opus 5 anthropic | ● | 44.2 | 11.7 | 49 | 2026-07 | try | |
| Muse Wonder meta | ● | 43.6 | 9.33 | 164 | 2026-10 | try | |
| Muse Spark 1.3 meta | ● | 36.6 | 2.25 | 146 | 2026-09 | try | |
| Claude Haiku 5.5 anthropic | ● | 36.5 | 0.233 | 165 | 2026-10 | try | |
| Grok 4.7 x-ai | ● | 35.5 | 3.33 | 55 | 2026-09 | try | |
| Hy4 tencent | ● | 34.6 | 0.695 | api price | 2026-11 | try | |
| Qwen3.8 Max (0902) qwen | ● | 33.7 | 3.33 | 2026-09 | try | ||
| GPT-5.6 Terra openai | ● | 32.3 | 5.33 | 95 | 2026-07 | try | |
| GLM 5.3 z-ai | ● | 31.7 | 0.108 | 72 | 6× B200 753B, 35B active | 2026-08 | try |
| Grok 4.6 x-ai | ● | 31 | 3.33 | 63 | 2026-08 | try | |
| Gemini 3.8 Flash google | ● | 30.8 | 1.93 | 136 | 2026-09 | try | |
| MiMo-V2.6-Pro xiaomi | ● | 29.9 | 0.0983 | 41 | 7× B200 1024B, 32B active | 2026-09 | try |
| MiniMax M3.1 Flash minimax | ● | 29.1 | 0.0317 | 152 | 2× B200 220B, 10B active | 2026-10 | try |
| DeepSeek V4.1 Flash 1017 deepseek | ● | 28.6 | 0.0549 | 155 | 6× B200 763B, 18B active | 2026-10 | try |
| Step 5 Preview stepfun | ● | 28.5 | 0.0633 | 150 | 5× B200 604B, 21B active | 2026-10 | try |
| Solar Pro 4 upstage | ● | 28.2 | 0.3 | 58 | 2026-08 | try | |
| GPT-6 Luna openai | ● | 27.9 | 0.233 | 118 | 2026-09 | try | |
| GLM 5.3 Flash z-ai | ● | 27.8 | 0.0411 | 55 | 3× B200 321B, 13B active | 2026-08 | try |
| MiMo-V2.6-Flash xiaomi | ● | 26.7 | 0.0447 | 37 | 3× B200 311B, 15B active | 2026-09 | try |
| Mistral Large 4 mistralai | ● | 26.1 | 0.16 | 128 | 7× B200 1000B, 52B active | 2026-10 | try |
| Muse Spark Open meta | ● | 24.9 | 0.129 | 214 | 6× B200 816B, 42B active | 2026-10 | try |
| GPT-5.6 Luna openai | ● | 23.8 | 0.533 | 105 | 2026-07 | try | |
| DeepSeek V4 Pro 0813 deepseek | ● | 23.1 | 1.63 | 79 | api price | 2026-08 | try |
| Qwen3.8 27B qwen | ● | 20.1 | 0.0853 | 36 | 1× B200 28B | 2026-08 | try |
| MiniMax M3 minimax | ● | 18.9 | 0.0615 | 121 | 3× B200 427B, 20B active | 2026-05 | try |
| Beam 1.0 501B A23B reflection | ● | 17.7 | 0.0706 | 4× B200 501B, 23B active | 2026-10 | try | |
| Kimi K3 moonshotai | ● | 17.4 | 5.46 | 50 | api price | 2026-07 | try |
| Nemotron 3.5 Super VL nvidia | ● | 13.6 | 0.267 | 266 | api price | 2026-10 | try |
| Inkling Small thinkingmachines | ● | 12.9 | 0.0287 | 81 | 2× B200 266B, 9B active | 2026-07 | try |
| DeepSeek V4 Pro 0423 deepseek | ● | 12.6 | 0.944 | api price | 2026-04 | try | |
| Inkling thinkingmachines | ● | 12.5 | 0.103 | 139 | 7× B200 952B, 33B active | 2026-07 | try |
| Nemotron 3 Ultra nvidia | ● | 11.8 | 0.169 | 4× B200 550B, 55B active | 2026-06 | try | |
| Gemini 3.5 Flash Lite google | ● | 11.7 | 1.03 | 392 | 2026-07 | try | |
| Claude Haiku 4.5 anthropic | ● | 7.4 | 2.33 | 2025-10 | try | ||
| Muse Glimmer 30B meta | ● | 6.3 | 0.0914 | 1× B200 30B | 2026-08 | try | |
| Command A+ cohere | ● | 6.2 | 0.7 | 247 | 2026-09 | try | |
| Gemini 3.1 Flash Lite Preview google | ● | 6.2 | 0.667 | 2026-03 | try | ||
| Nemotron 3.5 Lightning nvidia | ● | 5.5 | 0.00921 | 293 | 1× B200 30B, 3B active | 2026-08 | try |
| Granite 4.2 8B ibm-granite | ● | 5.1 | 0.027 | 124 | 1× B200 9B | 2026-08 | try |
| Mistral Medium 3.5 mistralai | ● | 4.6 | 3.5 | 172 | 2026-04 | try | |
| Gemma 4 31B google | ● | 4.4 | 0.096 | 35 | 1× B200 31B | 2026-04 | try |
| Mistral Small 4 mistralai | ● | 4.3 | 0.0172 | 160 | 1× B200 119B, 6B active | 2026-03 | try |
| North Mini Code cohere | ● | 3.4 | 0.00878 | 99 | 1× B200 30B, 3B active | 2026-06 | try |
A card is written off over 3 years at 50% utilization, plus electricity at $0.2/kWh, times 1.3 for the box around it. The cheapest card (at most 8 per box) whose memory holds the fp8 weights wins. An output token costs card-hours ÷ (peak TFLOPS × 0.2 ÷ 2 × active parameters), an input token the same at 0.6. No margin, no idle box, no engineer: a floor, not a quote. Parameter counts come from Hugging Face; for mixture-of-experts models the active share is estimated.
| card | $ | GB | TFLOPS | W | $ / h |
|---|---|---|---|---|---|
| RTX 5090 | 2000 | 32 | 210 | 575 | 0.50 |
| H100 | 30000 | 80 | 989 | 700 | 3.33 |
| B200 | 40000 | 180 | 2250 | 1000 | 4.48 |