Firefly Launches IP67 AI Cameras with 3 to 6 TOPS
Firefly announced two IP67-rated AI cameras, the CQ38W-1126B and CQ38W-3576, using Rockchip SoCs with 3 TOPS and 6 TOPS NPUs respectively. The cameras come in 3MP or 5MP sensor options and commercial, industrial (J-suffix), and automotive (M-suffix on the 3576) variants. The CQ38W-1126B uses the RV1126B SoC and targets small multimodal models, while the CQ38W-3576 uses the RK3576 octa-core SoC and can run heavier workloads including YOLO and compact LLMs. Both models move from PoE to a 12V DC input and add an RS485 interface, trading PoE convenience for higher compute and automotive-grade configurations. The devices aim at commercial, industrial, and automotive edge-AI deployments that require rugged enclosures and on-device inference.
What happened
Firefly released two new IP67-rated AI cameras, the CQ38W-1126B and CQ38W-3576, built around Rockchip silicon with 3 TOPS and 6 TOPS NPUs. The entry model uses the RV1126B SoC and the higher-end model uses the RK3576, available in commercial, industrial (J), and automotive (M) variants. Sensor choices are 3MP or 5MP, and video capabilities include up to 4K decoding and encoding depending on SoC.
Technical details
The product choices reflect clear SoC differentiation and real-world edge-AI design tradeoffs. The CQ38W-1126B uses RV1126B, a quad-core Cortex-A53 CPU up to 1.6 GHz, a 3 TOPS NPU (INT8) and multi-framework support (TensorFlow, ONNX, PyTorch, Caffe). The CQ38W-3576 uses RK3576, an octa-core CPU cluster including Cortex-A72 cores, a 6 TOPS NPU with INT4/INT8/INT16/BF16/TF32 support, and advanced VPU decoding including AV1 and 8Kp30. Key platform characteristics:
- •CPU/GPU/NPU balance: Cortex-A53 quad-core with 3 TOPS vs octa-core (A72/A53) with 6 TOPS
- •Memory options: 1-4GB LPDDR4 for 1126B, 4-16GB options for 3576
- •Interfaces: switched from PoE to 12V DC power, added RS485 serial link
- •Video: hardware decode/encode up to 4Kp60 to 8Kp30 depending on SoC
Context and significance
Edge camera vendors are moving NPUs from sub-1 TOPS into multi-TOPS territory so devices can run modern vision models and compact LLMs locally. The 3 TOPS part targets typical object detection and multimodal micromodels, while 6 TOPS and the RK3576 give headroom for YOLO-class detectors and small LLMs used for on-device inference or preprocessing before sending embeddings. The addition of an automotive-grade M-suffix and industrial J-suffix shows Firefly positioning the platform for harsher environments beyond retail and office surveillance.
What to watch
Evaluate whether the loss of PoE reduces deployment flexibility, and test model runtime and thermal behavior on the RK3576 under sustained workloads. Also watch software tooling and SDK support for model conversion and quantization to INT8/INT4 on these Rockchip NPUs.
Scoring Rationale
The release matters to practitioners deploying edge vision and automotive AI because it pushes NPUs into the 3-6 TOPS range in a rugged form factor. It is a solid hardware refresh rather than a paradigm shift, with tradeoffs like loss of PoE limiting some deployments.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problemsStep-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.
