# VLM Inference Latency on Qualcomm Snapdragon 8 Gen 3 SoC llama.cpp quantization

Source: [VLM Inference Latency on Qualcomm Snapdragon 8 Gen 3 SoC llama.cpp quantization](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T06:08:00.967Z

Evaluation of Vision-Language Model (VLM) inference performance, specifically measuring latency (ms/patch) and throughput (tokens/s) when running quantized models via llama.cpp on the Qualcomm Snapdragon 8 Gen 3 SoC.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | VLM Inference Latency |
| Dataset (including any stated split) | Qualcomm Snapdragon 8 Gen 3 SoC llama.cpp quantization |
| Metric | VE Latency (ms/patch) |
| Unit | TIME |
| Direction | Lower is better |
| Results loaded | 5 of 5 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [VE Latency (ms/patch)](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization.md?metric=VE%20Latency%20(ms%2Fpatch)>): Lower is better; unit: TIME.

- [Sample Throughput (tokens/s)](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization.md?metric=Sample%20Throughput%20(tokens%2Fs)>): Higher is better; unit: SCALAR.

- [Prompt Evaluation Throughput (tokens/s)](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization.md?metric=Prompt%20Evaluation%20Throughput%20(tokens%2Fs)>): Higher is better; unit: SCALAR.

- [Generation Throughput (tokens/s)](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization.md?metric=Generation%20Throughput%20(tokens%2Fs)>): Higher is better; unit: SCALAR.

- [Total Inference Time (s)](<https://www.sota2.com/research/sota/vlm-inference-latency-on-qualcomm-snapdragon-8-gen-3-soc-llama-cpp-quantization.md?metric=Total%20Inference%20Time%20(s)>): Lower is better; unit: TIME.

## Recorded results

| Method | VE Latency (ms/patch) | Result date | Report paper |
| --- | --- | --- | --- |
| MobileVLM-336 | 6.82 | Not recorded | 68fa71a71e209131cea52555 |
| LLaVA-v1.5-336 | 7.77 | Not recorded | 68fa71a71e209131cea52555 |
| LLaVA-v1.5-336 | 7.98 | Not recorded | 68fa71a71e209131cea52555 |
| LLaVA-v1.5-336 | 8.23 | Not recorded | 68fa71a71e209131cea52555 |
| MobileVLM-336 | 8.43 | Not recorded | 68fa71a71e209131cea52555 |

### MobileVLM-336

VE Latency (ms/patch): 6.82. Result ID: 6991f3df3a6c3148490ba8c0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Implementation](<https://github.com/Meituan-AutoML/MobileVLM>)

Recorded evaluation configuration: {"Language Model":"MobileLLaMA 1.4B","Size (GB)":"1.40","Quantization":"4-bit","Input Resolution":"336x336"}. This record does not assert independent reproduction.

### LLaVA-v1.5-336

VE Latency (ms/patch): 7.77. Result ID: 6991f3df3a6c3148490ba8be.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732e1e209131cea5af59.md>)

[Implementation](<https://llava-vl.github.io/>)

Recorded evaluation configuration: {"Language Model":"TinyLLaMA 1B","Size (GB)":"1.18","Quantization":"4-bit","Input Resolution":"336x336"}. This record does not assert independent reproduction.

### LLaVA-v1.5-336

VE Latency (ms/patch): 7.98. Result ID: 6991f3df3a6c3148490ba8bd.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732e1e209131cea5af59.md>)

[Implementation](<https://llava-vl.github.io/>)

Recorded evaluation configuration: {"Language Model":"OpenLLaMA 3B","Size (GB)":"2.88","Quantization":"4-bit","Input Resolution":"336x336"}. This record does not assert independent reproduction.

### LLaVA-v1.5-336

VE Latency (ms/patch): 8.23. Result ID: 6991f3df3a6c3148490ba8bc.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732e1e209131cea5af59.md>)

[Implementation](<https://llava-vl.github.io/>)

Recorded evaluation configuration: {"Language Model":"Vicuna 7B","Size (GB)":"4.70","Quantization":"4-bit","Input Resolution":"336x336"}. This record does not assert independent reproduction.

### MobileVLM-336

VE Latency (ms/patch): 8.43. Result ID: 6991f3df3a6c3148490ba8bf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71a71e209131cea52555.md>)

[Implementation](<https://github.com/Meituan-AutoML/MobileVLM>)

Recorded evaluation configuration: {"Language Model":"MobileLLaMA 2.7B","Size (GB)":"2.14","Quantization":"4-bit","Input Resolution":"336x336"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
