Raspberry Pi AI HAT+ 2 (Hailo-10H): Llama-3.2-1B compiled-model TPS (vendor table)
9.89 tokens/s (Hailo compiled-model TPS table, single-stream)
Fields
id
benchmarks/comm-hailo10h-pi-ai-hat2-llama32-1b-tpsMetric
decode_tok_s
Value
9.89
Unit
tokens/s (Hailo compiled-model TPS table, single-stream)
Model
Llama-3.2-1B-Instruct
Methodology
VENDOR-MEASURED. Hailo's compiled-model TPS table (hailo-ai/hailo_model_zoo_genai -> Hailo-Ollama, a real C++ Ollama-compatible server on HailoRT). The compiled model set tops out at <=2B parameters - THIS IS THE KEY FINDING: despite the 'GenAI' badge and 8 GB onboard RAM, there is NO 7B+ path. Full measured table: Llama-3.2-1B 9.89 TPS; Qwen2-1.5B 8.06; Qwen2.5-Coder-1.5B 8.13; DeepSeek-R1-Distill-1.5B 7.96; Qwen2.5-1.5B 7.35; Qwen3-1.7B 4.78; Qwen2-VL-2B 7.04.
Measured by
vendor
Reproducible
yes
notes
$200 official (+ ~$60 Pi 5). Hailo-10H, 40 TOPS INT4, 8 GB onboard RAM (model lives on the HAT, not the Pi). A real product with real measured tok/s and a smooth Ollama experience - genuinely impressive for $200 IF 1-2B models are the target. For the 7B goal it is a TRAP. NOTE: the ORIGINAL AI HAT+ ($70/$110) is Hailo-8/8L - vision only, no LLM; only the AI HAT+ 2 has Hailo-10H.
Notes
$200 official (+ ~$60 Pi 5). Hailo-10H, 40 TOPS INT4, 8 GB onboard RAM (model lives on the HAT, not the Pi). A real product with real measured tok/s and a smooth Ollama experience - genuinely impressive for $200 IF 1-2B models are the target. For the 7B goal it is a TRAP. NOTE: the ORIGINAL AI HAT+ ($70/$110) is Hailo-8/8L - vision only, no LLM; only the AI HAT+ 2 has Hailo-10H.