跳到主要内容

ModelZoo

ModelZoo内模型数据定期更新,模型应用示例请参考ai-sdk

基础模型

  • K1
  • 推理引擎版本: spacemit-ort-2.0.6
  • OS:bianbu-3.0
  • date:2026-7-27
  • K3
  • 推理引擎版本: v2.0.6
  • OS:bianbu-4.0rc3
  • date:2026-7-27

测试方式

# 进入spacemit-ort库路径
# cd {spacemit_ort_lib}/
export LD_LIBRARY_PATH=./lib/

# 调整为自己的${model_path}(模型文件路径),${num of cores}(选择跑几个核心)
./bin/onnxruntime_perf_test ${model_path} -e spacemit -r 10 -x 1 -S 1 -s -c 1 -i "SPACEMIT_EP_INTRA_THREAD_NUM|${num of cores}" -I

# 输出信息如下
using SpaceMITExecutionProvider
setting SPACEMIT_EP_INTRA_THREAD_NUM : 4
Setting intra_op_num_threads to 1
Session creation time cost: 0.169475 s
First inference time cost: 109 ms
Total inference time cost: 0.0727021 s
Total inference requests: 10
Average inference time cost total: 7.270205 ms
Total inference run time: 0.0727619 s
Number of inferences per second: 137.435
Avg CPU usage: 62 %
Peak working set size: 91336704 bytes
Avg CPU usage:62
Peak working set size:91336704
Runs:10
Min Latency: 0.00720383 s
Max Latency: 0.00730163 s
P50 Latency: 0.00727787 s
P90 Latency: 0.00730163 s
P95 Latency: 0.00730163 s
P99 Latency: 0.00730163 s
P999 Latency: 0.00730163 s

# Average inference time cost total即单帧推理耗时

resnet

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
resnet18int8224x22440.6822.1512.99
resnet50int8224x22495.3652.8932.29
resnet50fp16224x224674.48348.92213.96
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
resnet18int8224x2248.004.742.902.07
resnet50int8224x22421.0012.217.725.47
resnet50.batch4int8224x22477.9942.7524.6215.38
resnet50fp16224x22437.6421.9514.8911.36

mobilenet

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
mobilenet_v1int8224x22428.9915.179.14
mobilenet_v2int8224x22428.8317.8611.81
mobilenet_v3_smallfp16224x22425.8215.529.99
mobilenet_v3_largefp16224x22465.7939.7925.28
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
mobilenet_v1int8224x22412.677.213.912.35
mobilenet_v2int8224x22417.699.925.223.26
mobilenet_v3_smallfp16224x2248.715.143.232.75
mobilenet_v3_largefp16224x22416.599.675.954.52

efficientnet

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
efficientnet_v1_b0int8224x22480.7946.1029.91
efficientnet_v1_b1int8224x224115.9566.0942.71
efficientnet_v2_sint8224x224162.1191.7557.08
efficientnet_v1_b0fp16224x224136.7579.1953.62
efficientnet_v1_b1fp16224x224193.29112.9878.21
efficientnet_v2_sfp16224x224597.90323.13185.05
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
efficientnet_v1_b0int8224x22440.8422.5112.899.85
efficientnet_v1_b1int8224x22463.4634.4619.8014.75
efficientnet_v2_sint8224x22458.4133.4619.8013.53
efficientnet_v1_b0fp16224x22443.3523.6714.3010.91
efficientnet_v1_b1fp16224x22463.3735.1920.9615.63
efficientnet_v2_sfp16224x22482.8947.6829.7720.67

vit

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
vit_b_16int8224x224506.60328.01167.38
vit_b_16fp16224x2242478.261398.33765.65
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
vit_b_16int8224x224100.9056.6435.1923.76
vit_b_16fp16224x224138.4686.3762.1148.75

yolov5

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
yolov5nint8640x640242.92132.9179.59
yolov5sint8640x640465.30244.70142.46
yolov5mint8640x640959.66493.48273.47
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolov5nint8640x64043.7624.1114.369.57
yolov5sint8640x64073.4040.2023.9515.81
yolov5mint8640x640152.0381.9045.8828.97

yolov6

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
yolov6nint8640x640173.4293.2155.98
yolov6sint8640x640438.71224.39123.60
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolov6nint8640x64031.9018.0410.957.53
yolov6sint8640x64065.3735.9821.3313.42

yolov8

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
yolov8nint8640x640233.76123.0973.74
yolov8sint8640x640515.20273.62150.33
yolov8mint8640x6401050.21529.48298.47
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolov8nint8640x64041.4123.0813.859.51
yolov8sint8640x64074.5041.1924.8916.65
yolov8mint8640x640157.8085.6848.9032.31

yolov8-seg

  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolov8n-segint8640x64061.5933.5719.4012.75
yolov8s-segint8640x640103.2756.3333.2421.51
yolov8m-segint8640x640204.94109.9461.9339.61

yolov8-pose

  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolov8n-poseint8640x64046.4226.1816.1611.16
yolov8s-poseint8640x64081.6145.3327.8718.64
yolov8m-poseint8640x640165.6090.3051.7533.84

yolo12

  • K1
模型名typeshape1 Core/ms2 Core/ms4 Core/ms
yolo12nint8640x640365.20194.36118.23
yolo12sint8640x640795.61443.95246.50
yolo12mint8640x6401921.821004.42581.78
  • K3
模型名typeshape1 Core/ms2 Core/ms4 Core/ms8 Core/ms
yolo12nint8640x640107.5657.8532.5521.73
yolo12sint8640x640191.31102.0556.9036.24
yolo12mint8640x640378.55200.75110.2369.17

音频模型

  • K1
模型名type4 Core/rtf
melottsdyn_int80.984
sensevoicedyn_int8---
  • K3
模型名type4 Core/rtf8 Core/rtf
melottsdyn_int80.530---
sensevoicedyn_int80.11240.1380

大模型

  • K3
  • llama.cpp版本:0.1.1
  • OS:bianbu-4.0rc3
  • date:2026-5-26

测试方式

# 进入spacemit-llama.cpp库路径
# cd {spacemit-llama.cpp}/
export LD_LIBRARY_PATH=./lib/

# 调整为自己的${model_path}(模型文件路径),${num of cores}(选择跑几个核心)
./bin/llama-bench -m ${model_path} -t ${num of cores} -p 128 -n 128 -mmp 0 -fa 1 -ub 128

# 输出信息如下
CPU_RISCV64_SPACEMIT: tcm is available, blk_size: 393216, blk_num: 8, is_fake_tcm: 0
CPU_RISCV64_SPACEMIT: num_cores: 16, num_perfer_cores: 8, perfer_core_arch_id: a064, exclude_main_thread: 0, use_ime1: 0, use_ime2: 1, mem_backend: HPAGE, cpu_mask: ff00, aicpu_id_offset: 8
CPU_RISCV64_SPACEMIT: alloc_chunk: open(/dev/tcm_sync_mem) failed, errno=2
CPU_RISCV64_SPACEMIT: failed to allocate init_barrier from shared mem, falling back to heap
| model | size | params | backend | threads | n_ubatch | fa | mmap | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | ---: | --------------: | -------------------: |
| qwen3 0.6B Q4_0 | 358.78 MiB | 596.05 M | CPU | 8 | 128 | 1 | 0 | pp128 | 499.75 ± 0.22 |
| qwen3 0.6B Q4_0 | 358.78 MiB | 596.05 M | CPU | 8 | 128 | 1 | 0 | tg128 | 53.35 ± 0.03 |

Qwen

  • K3
模型名量化类型PP128 (token/s)TG128 (token/s)PP1280 (token/s)TG1280 (token/s)
qwen3-0.6BQ4_0499.7553.35--
qwen3-1.7BQ4_0229.7923.11--
qwen3-4BQ4_076.4411.03--
qwen3-moe-30B-A3BQ4_055.6712.3244.0311.17
qwen3.5-0.8BQ4_0182.6929.33--
qwen3.5-2BQ4_1112.2216.15--

HunYuan

  • K3
模型名量化类型PP128 (token/s)TG128 (token/s)PP1280 (token/s)TG1280 (token/s)
HY-MT1.5-1.8BQ4_K_M157.8120.15--

Llama

  • K3
模型名量化类型PP128 (token/s)TG128 (token/s)PP1280 (token/s)TG1280 (token/s)
llama2-7BQ4_050.407.07--

多模态大模型

  • K3
  • llama.cpp版本:0.1.1
  • 推理引擎版本: v2.0.3
  • OS:bianbu-4.0rc3
  • date:2026-5-26

测试方式

以qwen3vlencoder为例

export LD_LIBRARY_PATH=./spacemit-llama.cpp/lib:./spacemit_ort/lib
export SPACEMIT_EP_DENSE_ACCURACY_LEVEL=1

llama-server -m qwen3vl-30b-text-q4_1.gguf --media-backend smt --smt-config-dir ./ -ctk f16 -ctv f16 -t 8 -c 1024 --host 0.0.0.0 --port 8080 --reasoning-budget 0 --reasoning off

详细参数含义见llama.cpp.md

VLM

  • K3
模型名图像规格LLM 8 Core + VisionEncoder 4 Core/msLLM 8 Core + VisionEncoder 8 Core/ms
fastvlm-0.5B512*512256.47164.50
Qwen3-VL-30B-A3B768*7687928.134753.55
Qwen3.5-0.8B384*384340.42245.61
Qwen3.5-2B384*384901.56794.03
Qwen3.5-4B384*384904.73798.71

ASR

  • K3
模型名量化方式线程配置RTF
Qwen3-ASR 0.6BQ4_0 + 动态量化 ONNXLLM 8 / AudioEncoder 40.169
Qwen3-ASR 1.7BQ4_0 + 动态量化 ONNXLLM 8 / AudioEncoder 40.329
Fun-ASR NanoQ4_K_M + 量化 ONNXLLM 4 / AudioEncoder 40.247
Gemma4 ASR E2BQ4_0 + 量化 ONNXllama-server 8 / AudioEncoder 默认0.689(转写)/ 0.578(翻译)

Qwen3-ASR 0.6B 和 1.7B 均使用 001_zh_daily_weather.wav004_zh_selling_sausages.wav 连续测试 3 轮,总音频时长 47.331 秒,总处理时间分别为 7.977 秒和 15.572 秒。Fun-ASR Nano 使用中文、英文、日文三条音频测试,单轮总音频 时长 28.558 秒,预热后处理时间 7.063 秒。Gemma4 ASR 中文转写使用 004_zh_selling_sausages.wav,音频时长 14.158 秒,处理时间 9.758 秒;翻译使用 日语、韩语、粤语样例连续测试 2 轮,总音频时长 34.104 秒,预热后处理时间 19.703 秒。首个动态 ONNX encoder session 初始化请求未计入。表中 RTF 为端到端 结果。由于测试集和任务不同,各模型的 RTF 不宜直接横向比较识别或翻译质量。