What happened
Firefly has launched the AIBOX-9075, an industrial Edge AI box built around the Qualcomm DragonWing IQ-9075 SoC, per CNX-Software (June 26, 2026). The platform delivers up to 200 TOPS at INT8 and ships with 36GB LPDDR5 ECC memory, 128GB UFS 2.2 storage, and an M.2 PCIe 4.0 x4 NVMe slot. Connectivity includes dual 2.5GbE with TSN, optional Wi-Fi 6 and Bluetooth 5.2, optional 4G/5G via mini PCIe, 8x GMSL2 camera inputs, HDMI 2.0, USB 3.0, CAN-FD, RS485, and opto-isolated digital I/O. The unit is housed in a rugged aluminum enclosure rated -40C to 85C for continuous industrial and outdoor deployments. The AIBOX-9075 is priced at $1,239 direct from the Firefly store, per CNX-Software.
Technical details
The IQ-9075 SoC pairs an octa-core Kryo Gen 6 (Cortex-A78C) cluster clocked up to 2.36 GHz with quad Cortex-R52 real-time cores and an Adreno 663 GPU delivering up to 1.2 TFLOPS FP32. AI acceleration is provided by Qualcomm's Hexagon Tensor Processor (quad HVX, dual HMX) rated at 200 TOPS INT8 sparse. Video support comes via the Adreno VPU 765, handling up to 1x 8Kp60 decode or 2x 4Kp60 concurrent encode/decode, alongside a Qualcomm Spectra 690 ISP for the 8-channel GMSL2 camera pipeline.
Software support
Firefly ships the AIBOX-9075 on Ubuntu and Yocto Linux via the Qualcomm Linux software stack, per CNX-Software. Supported AI frameworks include QNN, TensorFlow, TensorFlow Lite, PyTorch, ONNX, Keras, and Caffe. The device runs DeepSeek-R1, Gemma, Llama, Qwen, and Phi LLMs, and Stable Diffusion for image generation. Firefly states the Hexagon NPU delivers 22 tokens/s with Llama2-7B and 12 tokens/s with 13B-parameter models.
Industry context
Edge AI appliances combining multi-TOPS NPUs, ECC memory, and industrial I/O are an established and growing segment for private LLM, robotics, and machine-vision workloads. The IQ-9075 is also used by competing products such as Radxa's Fogwise AIRbox Q900 and Innodisk's EXEC-Q911. The AIBOX-9075 differentiates on 8-channel GMSL2 camera support, CAN-FD/RS485 integration, and TSN networking, which are practical advantages for vehicle-mounted or factory-floor deployments that reduce systems integration overhead.
What to watch
BSP and driver maturity for the Hexagon Tensor Processor in mainline ML frameworks matters more than peak TOPS for real deployment decisions. Practitioners should monitor thermal and power profiles under sustained inference, and compare the 22 tokens/s Llama2-7B claim against benchmarks on competing IQ-9075 platforms.
Key Points #
- 1Firefly's AIBOX-9075 uses the Qualcomm DragonWing IQ-9075 for 200 TOPS edge inference with 36GB ECC RAM and 8x GMSL2 camera inputs, priced at $1,239.
- 2Software stack covers Ubuntu/Yocto Linux, QNN, PyTorch, ONNX, and LLMs including DeepSeek-R1 and Llama, with a vendor-claimed 22 tokens/s on Llama2-7B.
- 3TSN networking, CAN-FD, RS485, and -40C to 85C thermals make this a turnkey option for robotics, factory-floor vision, and on-premise LLM deployments.
Scoring Rationale #
AIBOX-9075 is a well-specified industrial edge AI appliance with Qualcomm IQ-9075 200 TOPS, ruggedized I/O, and confirmed LLM support, relevant to edge inference and robotics practitioners. It is a single-vendor product announcement in an increasingly crowded IQ-9075 appliance segment alongside Fogwise AIRbox Q900 and Innodisk EXEC-Q911. Niche-relevant for practitioners but not broadly impactful.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.