anthonypjshaw/harrier-oss-v1-270m-qnn-htp-b1-s128-perf
Feature Extraction • Updated
Harrier-OSS on Snapdragon NPU
Harrier embeddings compiled for Snapdragon X Elite Hexagon NPU via Qualcomm AI Hub. Up to 92 emb/s on-device via ORT QNN HTP execution provider.
Note RECOMMENDED. 92.94 emb/s on X Elite HTP (10.8 ms/embedding) - 5x faster than [1,512]. 419 MB. Gemma3 270M, INT8 static, [1, 128] input. Truncates inputs >128 tokens (chunk longer inputs). Best default for embedding short content.
anthonypjshaw/harrier-oss-v1-270m-qnn-htp
Feature Extraction • UpdatedNote Long-context fallback. 18.37 emb/s on X Elite HTP (54 ms/embedding). 440 MB. Gemma3 270M, [1, 512] input. Same model weights as the s128 build but compiled to accept up to 512 tokens per call. Use when inputs exceed 128 tokens and chunking is undesirable.
anthonypjshaw/harrier-oss-v1-0.6b-qnn-htp
Feature Extraction • UpdatedNote Quality variant. 4.41 emb/s on X Elite HTP (227 ms/embedding). 730 MB. Qwen3 0.6B, [1, 512] input. Higher retrieval quality than the 270M builds but 4-20x slower. Use only when retrieval recall matters more than throughput.