AT 3000 AI MiniPC / palm-sized edge AI inference box (Series)

Specifications

• CPU: Changjiang 100 SOC (Domestic)Supplier claim
• NPU: Houmo M50 24GSupplier claim
• AI Compute: 370 INT8 (Supports 1x or 2x Houmo M50 24G NPU Expansion) TOPSSupplier claim
• RAM: 32GB LPDDR5x(Max. 64GB)Supplier claim
• ROM: 2×M.2 SSD 2280 NVMe (Max. 8TB)Supplier claim
• I/O: 1*HDMI® 1.4；1*DP 1.4；1*USB-C ；4*USB-A 3.2；2*RJ45 1G；WiFi 6 + BT 5.2Supplier claim
• Dimensions: 152*152*45.8mm mmSupplier claim
• Banner TOPS range: 210-370 TOPSSupplier claim
• Benchmark qwen3-30b-a3b (32K Context Length, MoE Architecture, 22.3GB VRAM Footprint), batch_size=1, GGUF quantized: Input/Output Length (Tokens) | Total Latency (s) | Generation Throughput (tok/s) | Overall Throughput (tok/s) | TTFT(ms) | TPOT(ms) | ITL(ms) ;; 128/128 / 5.23 / 24.49 / 48.98 / 314.44 / 38.67 / 38.37 ; 256/256 / 10.22 / 25.04 / 50.09 / 310.13 / 38.87 / 38.72 ; 512/512 / 20.35 / 25.16 / 50.33 / 552.71 / 38.73 / 38.66 ; 1024/1024 / 43.73 / 23.42 / 46.84 / 1,075.74 / 41.69 / 41.65 ; 2048/2048 / 91.61 / 22.36 / 44.71 / 2,189.43 / 43.68 / 43.66 ; 3072/3072 / 143.54 / 21.39 / 42.80 / 3,416.32 / 45.64 / 45.61 ; 4096/4096 / 197.60 / 20.73 / 41.46 / 4,703.48 / 47.11 / 47.09 ; 8192/8192 / 457.10 / 17.92 / 35.84 / 10,891.69 / 54.48 / 54.47 ; 16384/16384 / 1,158.22 / 14.15 / 28.29 / 27,277.62 / 69.03 / 69.03Supplier claim
• Benchmark qwen3-8b (32K Context Length, 6.33GB VRAM Footprint): Input/Output Length (Tokens) | Total Latency (s) | Generation Throughput (tok/s) | Overall Throughput (tok/s) | TTFT(ms) | TPOT(ms) | ITL(ms) ;; 128/128 / 6.95 / 18.41 / 36.82 / 184.53 / 53.29 / 52.87 ; 256/256 / 13.92 / 18.39 / 36.79 / 190.16 / 53.83 / 53.62 ; 512/512 / 28.23 / 18.14 / 36.27 / 337.36 / 54.59 / 54.48 ; 1024/1024 / 56.63 / 18.08 / 36.16 / 631.87 / 54.74 / 54.69 ; 2048/2048 / 118.09 / 17.34 / 34.68 / 1,291.83 / 57.06 / 57.03 ; 3072/3072 / 188.16 / 16.33 / 32.65 / 2,036.30 / 60.61 / 60.59 ; 4096/4096 / 260.26 / 15.74 / 31.48 / 2,841.30 / 62.86 / 62.85 ; 8192/8192 / 606.56 / 13.51 / 27.01 / 6,721.81 / 73.23 / 73.22 ; 16384/16384 / 1,566.99 / 10.46 / 20.91 / 17,661.82 / 94.57 / 94.56Supplier claim
• Benchmark qwen3-14b (16K Context Length, 10.1GB VRAM Footprint): Input/Output Length (Tokens) | Total Latency (s) | Generation Throughput (tok/s) | Overall Throughput (tok/s) | TTFT(ms) | TPOT(ms) | ITL(ms) ;; 128/128 / 11.86 / 10.79 / 21.59 / 289.76 / 91.08 / 90.37 ; 256/256 / 23.76 / 10.78 / 21.55 / 299.87 / 91.99 / 91.63 ; 512/512 / 47.48 / 10.78 / 21.57 / 555.13 / 91.82 / 91.64 ; 1024/1024 / 97.30 / 10.52 / 21.05 / 1,075.79 / 94.06 / 93.97 ; 2048/2048 / 199.70 / 10.26 / 20.51 / 2,273.13 / 96.45 / 96.40 ; 3072/3072 / 306.78 / 10.01 / 20.03 / 3,606.10 / 98.72 / 98.69 ; 4096/4096 / 422.42 / 9.70 / 19.39 / 5,082.04 / 101.91 / 101.89 ; 8192/8192 / 942.44 / 8.69 / 17.38 / 12,240.25 / 113.56 / 113.55Supplier claim