Axera has unveiled the AX8860, a 6-nanometer edge AI chip series aimed at running large models on factory floors, warehouses and robots rather than in distant data centers. The company presented the line at CIIF 2026 in Shanghai and framed it as a compute base for physical AI that must sense, decide and act on site.
According to Axera's official release, the AX8860 delivers up to 256 TOPS of mixed-precision compute with native FP4 support, up to 273 GB/s of memory bandwidth and as much as 128 GB of LPDDR5X. It supports Axera's AXLink multi-chip cascading and PCIe 5.0 for scaling, plus 64-channel video decode and 16-channel encode for multi-camera vision workloads.
Around the chip, Axera also showed the Yuanxi A256S module and related inference cards and servers. The software stack is meant to look familiar to cloud developers: the company says the series supports vLLM, SGLang and PyTorch and offers an OpenAI API-compatible service interface, with tools covering quantization, conversion, compilation, profiling and debugging.
Partner channel claims have drawn attention beyond the official brief. Hardware maker Sipeed, cited by Wccftech, said it measured about 25% faster prefill and 30% faster decode than an NVIDIA DGX Spark on a Qwen 27B run, and described pricing as roughly half that of a DGX Spark. Those figures are Sipeed's claims, not a third-party lab benchmark, and are not stated in Axera's own CIIF announcement.
Wccftech also reported module designs around 40 W, single- and four-chip PCIe cards, and a roadmap toward higher-TOPS follow-ons. Axera says four-chip cascading can reach about 1,024 TOPS. The company positions the line for smart manufacturing, logistics and robotics, arguing that inference should happen where events occur rather than round-tripping to the cloud.
SEE ALSO: Axera Debuts 5nm M9 ADAS SoC Series With Up to About 720 TOPS