Hussain-Nazary

Edge & Embedded Inference Hardware

For cameras, robots, kiosks, vehicles, drones, and on-device AI with no cloud connection. Selection is driven by power budget (watts) and model class (tiny vision nets vs on-device LLMs). All model picks lead with the newest releases (Aug 2026) — Ministral 3 is the edge-native LLM line, Voxtral the edge-native STT.

NVIDIA Jetson Thor

The flagship edge module for robotics / “physical AI”.

Spec	Value
AI compute	2,070 FP4 TFLOPS (~7.5× AGX Orin; ~3.5× efficiency)
Memory	128 GB LPDDR5x
Power	40–130 W (configurable)
Module	Jetson T5000; developer kit available
Software	JetPack, Isaac (robotics), CUDA

Best for: Autonomous robots, drones, industrial inspection, and on-device LLM/VLM agents — it’s a DGX-Spark-class compute in an embedded form factor.

Recommended AI models (newest first):

LLM/VLM on device: Ministral 3 3B/8B/14B (Dec 2025 — Mistral explicitly optimized it for Jetson) → 8B: 20–40 tok/s, 14B: 12–18 tok/s; GLM-4.7-Flash Q4 (30B-A3B, Jan 2026) → ~8–12 tok/s; Gemma 3 27B (Jul 2026, native multimodal) → ~10 tok/s — vision-language for robots.
Bigger (slow but fits 128 GB): Mistral Medium 3 (Jul 2026) → ~5–6 tok/s; gpt-oss-120B Q4 → ~4–5 tok/s.
Classic vision: YOLOv11/v12 (real-time at 100+ FPS), DeepLab, EfficientDet for detection/segmentation pipelines.
Speech: Voxtral Transcribe 2 (Feb 2026, on-device STT with diarization — beats Whisper on FLEURS), Kokoro TTS for robot voice.

NVIDIA Jetson AGX Orin (64 GB)

Spec	Value
AI compute	275 TOPS (INT8)
Memory	64 GB LPDDR5
Power	15–60 W
Price	~$2,000 (dev kit / module)

Best for: Production edge AI — the workhorse of industrial/retail/agriculture deployments.

Recommended AI models (newest first):

Vision: YOLOv11/v12 (real-time), DeepStream pipelines for multi-camera, PoseNet, OCR (PaddleOCR).
Language: Ministral 3 3B/8B (Dec 2025) → 8B: 10–25 tok/s; GLM-4.7-Flash Q4 (Jan 2026) → ~6–10 tok/s; Gemma 3 12B → 10–15 tok/s.
Audio: Voxtral Transcribe 2; F5-TTS (small).

NVIDIA Jetson Orin Nano Super ($249 dev kit)

Spec	Value
AI compute	67 TOPS (INT8) — “Super” mode
Memory	8 GB LPDDR5
Power	7–25 W
Price	~$249 (dev kit)

Best for: The budget entry point into edge AI — hobbyists, students, small products.

Recommended AI models (newest first):

Vision: YOLOv8n/s/v11n (real-time), MobileNet, EfficientNet-Lite, PoseNet, FaceNet.
Language (small only): Ministral 3 3B Q4 (Dec 2025) → 5–10 tok/s; Qwen3.5 sub-12B smallest variant; Gemma 4 sub-12B tiny.
Audio: Voxtral Transcribe 2 (tiny profile) or Whisper tiny/base/small.
Embeddings: bge-small-en — fast enough for on-device RAG.

Hailo-8 / Hailo-8L / Hailo-10

Ultra-low-power NPU accelerator cards (M.2 / USB / PCIe).

Spec	Hailo-8	Hailo-8L	Hailo-10
AI compute	26 TOPS (INT8)	13 TOPS	40 TOPS
Power	~2.5 W	~1.5 W	~2.5 W
Use	M.2 card, NVRs, cameras	Raspberry Pi AI Kit, laptops	2025+ — laptops, edge boxes

Best for: Adding cheap NPU acceleration to existing PCs, NVRs, and the Raspberry Pi 5. Great at vision; transformer support (including small LLMs) has improved steadily.

Recommended AI models (newest first):

Vision (primary): YOLOv11/v12, ResNet, MobileNet, EfficientDet, FaceNet, pose estimation — 30–200 FPS depending on model size.
Language (Hailo-8/10): Ministral 3 3B (edge-optimized), Gemma 4 / Qwen3.5 sub-12B smallest variants (via Hailo’s transformer support — more limited than GPU options).
Audio: Voxtral Transcribe 2 (tiny) or Whisper tiny/base.

Google Coral (Edge TPU)

Spec	Value
AI compute	~4 TOPS (INT8)
Power	~2 W
Form factors	USB accelerator, M.2, dev board
Price	~$25–150

Best for: The cheapest, most power-efficient way to run lightweight vision models continuously (battery-powered cameras, always-on sensing).

Recommended AI models:

MobileNetV2/V3, EfficientNet-Lite, SSD-MobileNet, PoseNet, DeepLab — all at 30–100 FPS on 2 W.
Not for: LLMs or anything beyond tiny vision nets — 4 TOPS is a hard ceiling.

Raspberry Pi 5 + AI Kit (Hailo-8L)

Spec	Value
AI compute	13 TOPS (Hailo-8L)
CPU	Quad-core Cortex-A76 (Pi 5)
RAM	4/8/16 GB
Price	~$80 (Pi) + ~$70 (AI Kit)

Best for: The ultimate hobbyist local-AI box — this repo’s run-llm-raspberry-pi.html covers it in depth.

Recommended AI models (newest first):

Vision: YOLOv11n (real-time), MobileNet, pose estimation.
Language: Ministral 3 3B (Dec 2025) → 2–5 tok/s (slow but real); Qwen3.5 sub-12B smallest variant; Voxtral Transcribe 2 for voice.
Practical use: voice assistants, smart-home agents, on-device photo classification, tiny RAG.

Quick decision table

Device	TOPS	Power	Price	Model class	Newest-model sweet spot
Jetson Thor	2,070 FP4	40–130 W	~$2K+ (module)	3–120B LLM/VLM	Ministral 3 14B, GLM-4.7-Flash, Gemma 3 27B
Jetson AGX Orin	275	15–60 W	~$2K	1–30B + vision	Ministral 3 8B, GLM-4.7-Flash
Orin Nano Super	67	7–25 W	$249	1–3B + vision	Ministral 3 3B
Hailo-8/8L/10	13–40	1.5–2.5 W	$30–100	Vision + small LLM	YOLOv11/v12, Ministral 3 3B
Coral Edge TPU	4	2 W	$25–150	Tiny vision	MobileNet/EfficientNet-Lite
Pi 5 + AI Kit	13	~10 W total	~$150	1–3B + vision	Ministral 3 3B, Voxtral

Bottom line: pick by power budget first (watts decide everything at the edge), then by model class. Jetson = real compute, Hailo/Coral = ultra-efficient vision, Pi + AI Kit = the $150 tinkerer’s dream — and on all of them, Ministral 3 and Voxtral are the newest models built for exactly this class of hardware.