ModelZoo

Model data in ModelZoo is updated regularly. For model application examples, refer to ai-sdk.

• ModelZoo
  • Base Models
    • Benchmark Method
    • ResNet
    • MobileNet
    • EfficientNet
    • ViT
    • YOLOv5
    • YOLOv6
    • YOLOv8
    • YOLOv8-Seg
    • YOLOv8-Pose
    • YOLOv12
    • Audio Models
  • Large Language Models
    • LLM Benchmark Method
    • Qwen
    • HunYuan
    • Llama
  • Multimodal Large Models
    • Multimodal Benchmark Method
    • VLM
    • ASR

Base Models

• K1

• Inference engine version: spacemit-ort-2.0.6
• OS: bianbu-3.0
• Date: 2026-7-27

• K3

• Inference engine version: v2.0.6
• OS: bianbu-4.0rc3
• Date: 2026-7-27

Benchmark Method

# Enter the spacemit-ort library path
# cd {spacemit_ort_lib}/
export LD_LIBRARY_PATH=./lib/

# Replace ${model_path} with your own model file path and ${num of cores} with the number of CPU cores to use
./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

# Example output:
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" indicates the single-frame inference latency.

ResNet

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
resnet18	int8	224x224	40.68	22.15	12.99
resnet50	int8	224x224	95.36	52.89	32.29
resnet50	fp16	224x224	674.48	348.92	213.96

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
resnet18	int8	224x224	8.00	4.74	2.90	2.07
resnet50	int8	224x224	21.00	12.21	7.72	5.47
resnet50.batch4	int8	224x224	77.99	42.75	24.62	15.38
resnet50	fp16	224x224	37.64	21.95	14.89	11.36

MobileNet

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
mobilenet_v1	int8	224x224	28.99	15.17	9.14
mobilenet_v2	int8	224x224	28.83	17.86	11.81
mobilenet_v3_small	fp16	224x224	25.82	15.52	9.99
mobilenet_v3_large	fp16	224x224	65.79	39.79	25.28

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
mobilenet_v1	int8	224x224	12.67	7.21	3.91	2.35
mobilenet_v2	int8	224x224	17.69	9.92	5.22	3.26
mobilenet_v3_small	fp16	224x224	8.71	5.14	3.23	2.75
mobilenet_v3_large	fp16	224x224	16.59	9.67	5.95	4.52

EfficientNet

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
efficientnet_v1_b0	int8	224x224	80.79	46.10	29.91
efficientnet_v1_b1	int8	224x224	115.95	66.09	42.71
efficientnet_v2_s	int8	224x224	162.11	91.75	57.08
efficientnet_v1_b0	fp16	224x224	136.75	79.19	53.62
efficientnet_v1_b1	fp16	224x224	193.29	112.98	78.21
efficientnet_v2_s	fp16	224x224	597.90	323.13	185.05

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
efficientnet_v1_b0	int8	224x224	40.84	22.51	12.89	9.85
efficientnet_v1_b1	int8	224x224	63.46	34.46	19.80	14.75
efficientnet_v2_s	int8	224x224	58.41	33.46	19.80	13.53
efficientnet_v1_b0	fp16	224x224	43.35	23.67	14.30	10.91
efficientnet_v1_b1	fp16	224x224	63.37	35.19	20.96	15.63
efficientnet_v2_s	fp16	224x224	82.89	47.68	29.77	20.67

ViT

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
vit_b_16	int8	224x224	506.60	328.01	167.38
vit_b_16	fp16	224x224	2478.26	1398.33	765.65

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
vit_b_16	int8	224x224	100.90	56.64	35.19	23.76
vit_b_16	fp16	224x224	138.46	86.37	62.11	48.75

YOLOv5

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
yolov5n	int8	640x640	242.92	132.91	79.59
yolov5s	int8	640x640	465.30	244.70	142.46
yolov5m	int8	640x640	959.66	493.48	273.47

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolov5n	int8	640x640	43.76	24.11	14.36	9.57
yolov5s	int8	640x640	73.40	40.20	23.95	15.81
yolov5m	int8	640x640	152.03	81.90	45.88	28.97

YOLOv6

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
yolov6n	int8	640x640	173.42	93.21	55.98
yolov6s	int8	640x640	438.71	224.39	123.60

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolov6n	int8	640x640	31.90	18.04	10.95	7.53
yolov6s	int8	640x640	65.37	35.98	21.33	13.42

YOLOv8

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
yolov8n	int8	640x640	233.76	123.09	73.74
yolov8s	int8	640x640	515.20	273.62	150.33
yolov8m	int8	640x640	1050.21	529.48	298.47

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolov8n	int8	640x640	41.41	23.08	13.85	9.51
yolov8s	int8	640x640	74.50	41.19	24.89	16.65
yolov8m	int8	640x640	157.80	85.68	48.90	32.31

YOLOv8-Seg

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolov8n-seg	int8	640x640	61.59	33.57	19.40	12.75
yolov8s-seg	int8	640x640	103.27	56.33	33.24	21.51
yolov8m-seg	int8	640x640	204.94	109.94	61.93	39.61

YOLOv8-Pose

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolov8n-pose	int8	640x640	46.42	26.18	16.16	11.16
yolov8s-pose	int8	640x640	81.61	45.33	27.87	18.64
yolov8m-pose	int8	640x640	165.60	90.30	51.75	33.84

YOLO12

• K1

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms
yolo12n	int8	640x640	365.20	194.36	118.23
yolo12s	int8	640x640	795.61	443.95	246.50
yolo12m	int8	640x640	1921.82	1004.42	581.78

• K3

Model	Type	Shape	1 Core/ms	2 Core/ms	4 Core/ms	8 Core/ms
yolo12n	int8	640x640	107.56	57.85	32.55	21.73
yolo12s	int8	640x640	191.31	102.05	56.90	36.24
yolo12m	int8	640x640	378.55	200.75	110.23	69.17

Audio Models

• K1

Model	Type	4 Core/rtf
melotts	dyn_int8	0.984
sensevoice	dyn_int8	---

• K3

Model	Type	4 Core/rtf	8 Core/rtf
melotts	dyn_int8	0.530	---
sensevoice	dyn_int8	0.1124	0.1380

Large Language Models

• K3

  • llama.cpp version: 0.1.1
  • OS: bianbu-4.0rc3
  • Date: 2026-5-26

LLM Benchmark Method

# Enter the spacemit-llama.cpp library path
# cd {spacemit-llama.cpp}/
export LD_LIBRARY_PATH=./lib/

# Replace ${model_path} with your own model file path and ${num of cores} with the number of CPU cores to use
./bin/llama-bench -m ${model_path} -t ${num of cores} -p 128 -n 128 -mmp 0 -fa 1 -ub 128

# Example output:
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

Model	Quantization	PP128 (token/s)	TG128 (token/s)	PP1280 (token/s)	TG1280 (token/s)
qwen3-0.6B	Q4_0	499.75	53.35	-	-
qwen3-1.7B	Q4_0	229.79	23.11	-	-
qwen3-4B	Q4_0	76.44	11.03	-	-
qwen3-moe-30B-A3B	Q4_0	55.67	12.32	44.03	11.17
qwen3.5-0.8B	Q4_0	182.69	29.33	-	-
qwen3.5-2B	Q4_1	112.22	16.15	-	-

HunYuan

• K3

Model	Quantization	PP128 (token/s)	TG128 (token/s)	PP1280 (token/s)	TG1280 (token/s)
HY-MT1.5-1.8B	Q4_K_M	157.81	20.15	-	-

Llama

• K3

Model	Quantization	PP128 (token/s)	TG128 (token/s)	PP1280 (token/s)	TG1280 (token/s)
llama2-7B	Q4_0	50.40	7.07	-	-

Multimodal Large Models

• K3

  • llama.cpp version: 0.1.1
  • Inference engine version: v2.0.3
  • OS: bianbu-4.0rc3
  • Date: 2026-5-26

Multimodal Benchmark Method

Using qwen3vlencoder as an example:

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

For detailed parameter descriptions, see llama.cpp.

VLM

• K3

Model	Image Size	LLM 8 Core + VisionEncoder 4 Core/ms	LLM 8 Core + VisionEncoder 8 Core/ms
fastvlm-0.5B	512*512	256.47	164.50
Qwen3-VL-30B-A3B	768*768	7928.13	4753.55
Qwen3.5-0.8B	384*384	340.42	245.61
Qwen3.5-2B	384*384	901.56	794.03
Qwen3.5-4B	384*384	904.73	798.71

ASR

• K3

Model	Quantization	Thread Configuration	RTF
Qwen3-ASR 0.6B	Q4_0 + dynamic ONNX	LLM 8 / AudioEncoder 4	0.169
Qwen3-ASR 1.7B	Q4_0 + dynamic ONNX	LLM 8 / AudioEncoder 4	0.329
Fun-ASR Nano	Q4_K_M + quantized ONNX	LLM 4 / AudioEncoder 4	0.247
Gemma4 ASR E2B	Q4_0 + quantized ONNX	llama-server 8 / AudioEncoder default	0.689 (transcription) / 0.578 (translation)

Qwen3-ASR 0.6B and 1.7B were each measured for three rounds with 001_zh_daily_weather.wav and 004_zh_selling_sausages.wav: 47.331 seconds of audio and 7.977 and 15.572 seconds of total processing time, respectively. Fun-ASR Nano was measured with one Chinese, one English, and one Japanese sample: 28.558 seconds of audio per round and 7.063 seconds of processing time after warmup. Gemma4 ASR transcription used 004_zh_selling_sausages.wav: 14.158 seconds of audio and 9.758 seconds of processing time. Its translation measurement used two rounds of Japanese, Korean, and Cantonese speech-to-English samples: 34.104 seconds of audio and 19.703 seconds of processing time after warmup. The first request that initialized the dynamic ONNX encoder session was excluded. The reported RTF values are end-to-end results. Because the test sets and tasks differ, these RTF values should not be used to directly compare recognition or translation quality.