秒推算力节点
承接推理请求的机器。首字延迟、单流速度、吞吐与缓存命中是最近一小时的实测值，不是标称参数。
24
集群
一个运营方的一种模型算一个
134
在线机器
231
GPU
9
独立运营方
共 12 个运营方
GPU 构成RTX 5090 · 186其他 · 41H200 · 4
| 17× | GPU provider | kimi-k3 | 0.56s | 91 tps | 2,619 tps | 92% |
| 4× | GPU provider独立 | kimi-k3 | 0.54s | 91 tps | 559 tps | 87% |
| 2× | 8x NVIDIA B300 SXM6 AC独立 | kimi-k3 | 0.37s | 154 tps | 456 tps | 90% |
| 1× | 8x NVIDIA B300 SXM6 AC独立 | kimi-k3 | 0.49s | 90 tps | 399 tps | 92% |
| 1× | 8x NVIDIA B300 SXM6 AC独立 | kimi-k3 | 0.54s | 126 tps | 313 tps | 93% |
| 1× | 8x NVIDIA B300 SXM6 AC独立 | kimi-k3 | 0.46s | 154 tps | 246 tps | 91% |
| 1× | 1x NVIDIA GB10独立 | qwen3.6-35b-a3b | 0.84s | 21 tps | 151 tps | 49% |
| 5× | 8x RTX 5090 (32GB) 另有 1 种配置 | qwen3.8-27b | — | — | — | — |
| 2× | 32x RTX 5090 (32GB) | glm-5.3 | — | — | — | — |
| 2× | 32x RTX 5090 (32GB) | glm-5.2 | — | — | — | — |
| 1× | 16x RTX 5090 (32GB) | glm-5.3-flash | — | — | — | — |
| 1× | 2x NVIDIA GeForce RTX 5090独立 | qwen3.6-35b-a3b | — | — | — | — |
| 11× | GPU provider | glm-5.2 | — | — | — | — |
| 50× | GPU provider | deepseek-v4.1-flash | — | — | — | — |
| 14× | GPU provider | kimi-k3 | — | — | — | — |
| 4× | GPU provider | glm-5.3-flash | — | — | — | — |
| 1× | GPU provider | qwen3.8-27b | — | — | — | — |
| 4× | GPU provider | glm-5.3 | — | — | — | — |
| 5× | GPU provider | deepseek-v4-flash-0731 | — | — | — | — |
| 3× | GPU provider | qwen3.6-35b-a3b | — | — | — | — |
| 1× | 4x NVIDIA H200独立 | deepseek-v4.1-flash | — | — | — | — |
| 1× | PD: 2x 8xRTX5090 PP8 prefill + 2x TP4 H100 decode独立 | glm-5.3-flash | — | — | — | — |
| 1× | PD: 2x 8xRTX PRO 6000 PP8 prefill + 4x B200 TP4 decode独立 | glm-5.3 | — | — | — | — |
| 1× | PD: 3x 8xRTX5090 PP8 encoder-only prefill + 4x B200 DP4 decode独立 | deepseek-v4.1-flash | — | — | — | — |