Hailo-10H (Pi AI HAT+ 2), HailoRT/Ollama: Qwen3-1.7B compiled-model TPS (vendor table)
4.78 tokens/sec (Hailo compiled-model TPS table, single-stream)
Fields
id
benchmarks/comm-hailo10h-qwen3-1p7b-tpsMetric
decode_tok_s
Value
4.78
Unit
tokens/sec (Hailo compiled-model TPS table, single-stream)
Model
Qwen3-1.7B (Hailo-compiled)
Methodology
VENDOR-MEASURED compiled-model TPS table (hailo-ai/hailo_model_zoo_genai MODELS.rst, cited by research-exotic.md). Qwen3-1.7B on the Hailo-10H (Raspberry Pi AI HAT+ 2, 40 TOPS INT4, 8GB onboard): 4.78 TPS. This is the *largest* model class the compiled set tops out at - the model zoo's GenAI list ends at <=2B models, which is the whole reason the AI HAT+ 2 is a trap for the 7B goal.
Measured by
vendor
Reproducible
yes
notes
Gap-fill for the Oct-14 Phase-A ingest: research-exotic.md gave the full Hailo TPS table (Llama-3.2-1B 9.89, Qwen2-1.5B 8.06, Qwen2.5-1.5B 7.35, DeepSeek-R1-Distill-1.5B 7.96, Qwen2-VL-2B 7.04) but only the 1B and 1.5B rows had records. This row is the ceiling data point - the biggest model Hailo supports is also its slowest, which quantifies the trap.
Notes
Gap-fill for the Oct-14 Phase-A ingest: research-exotic.md gave the full Hailo TPS table (Llama-3.2-1B 9.89, Qwen2-1.5B 8.06, Qwen2.5-1.5B 7.35, DeepSeek-R1-Distill-1.5B 7.96, Qwen2-VL-2B 7.04) but only the 1B and 1.5B rows had records. This row is the ceiling data point - the biggest model Hailo supports is also its slowest, which quantifies the trap.