Axera Launches AX8860 High-Compute Chip, Targeting On-Site Inference for Edge and Device-Side Large Models

AI chip company Axera appeared at CIIF 2026, the China International Industrial Fair, under the theme "Smart Manufacturing Without Boundaries, Computing Power Delivered," laying out its AI computing power foundation and product lineup for the real physical world and formally launching its AX8860 high-compute series. Built specifically to run large models at the edge and on devices, the AX8860 series marks a shift in edge AI from technology validation into a new phase of end-to-end, business-value-driven deployment.

Four Exhibition Zones Show How AI Computing Power Gets Deployed

At this year's fair, Axera set up four zones: core computing power and technology foundation, industry solutions, software architecture and toolchains, and ecosystem and value co-creation. Together they presented the company's full-scenario computing offering for physical AI. With the AX8860 at the center, Axera showed a product line running from chips and core modules to inference cards and inference servers, covering everything from the computing foundation to application deployment for smart manufacturing.

In the ecosystem and value co-creation zone, Axera partner Yuanli Lingji showed its in-house DM0.5 general-purpose embodied brain, demonstrating three tasks: connector plug-and-unplug quality inspection, Insert Anything general industrial insertion, and flexible cutting with a dexterous hand. The demos pointed to the potential of embodied intelligence in electronics assembly, precision manufacturing, industrial quality inspection and flexible machining. DM0.5, the intelligence layer behind the demonstration, ran its model inference on Axera's high-compute platform, providing a single computing backbone for all three tasks.

AX8860: An Edge Computing Foundation Built for On-Site Inference

Axera argues that as large models, agents and embodied intelligence keep advancing, AI is moving from the information world into the physical world, making edge and device-side computing power a key requirement for real-time response in smart manufacturing. On demanding factory floors, the company says, a complete loop of on-site intelligence has to meet four conditions at once: it has to see, compute accurately, react fast, and execute reliably. Only when equipment on site closes the loop from sensing to judgment to execution can intelligence respond at the moment an event occurs, turning edge AI into a productive force behind industrial upgrading.

Designed to bring large models to the edge and device side, the AX8860 is not built around peak specifications alone. It targets practical needs such as model deployment, on-site inference and system coordination. Its core specifications:

  • Computing power and memory: 256 TOPS of mixed-precision computing power with native FP4 support, 273 GB/s of memory bandwidth, and 128 GB of LPDDR5X memory.
  • Interconnect and scaling: support for AXLink cascading and PCIe 5.0 high-speed interconnect, laying the groundwork for multi-chip coordination and elastic scaling of computing power.
  • Video: 64 channels of video decoding and 16 channels of video encoding, suited to multi-channel visual sensing and video analysis.
  • Process: 6 nm, balancing computing power, energy efficiency and integration.

Axera says this configuration meets the demand for high edge-side computing power in smart manufacturing, warehousing and logistics, and robotics. Alongside the AX8860, the company released the Yuanxi A256S module and showed its Yuanxi inference cards and inference servers. The A256S uses a modular design the company describes as compact, powerful and quick to deploy, supporting LLM, VLM and Omni model applications and offering more flexible ways to pair core computing power with application hardware.

Qiu Xiaoshen, Axera's founder and chair, said at the launch: "We want to be a stable, long-term computing foundation for our customers, so companies can focus on business innovation instead of worrying about migrating their compute. Axera is committed to bringing intelligence to every site, and the AX8860 high-compute series delivers on that promise."

A Full Product Lineup to End the "Which Compute Tier?" Dilemma

With the AX8860 series now official, Axera has extended its AI chip map again. The company has built out a full product matrix spanning low to very high computing power to cover differing compute needs.

From chips and core modules to inference cards and inference servers, Axera is offering a continuous product range with no gaps, letting customers scale computing power along a single technology roadmap and cut the cost of repeated vendor selection and system migration. As the company stressed at the launch, customers can pick the compute tier that fits their current stage of business and keep scaling along the same technology path, rather than facing the "which compute tier?" question over and over.

On the software side, the AX8860 series comes with a full toolchain and open-source software for developers and industry users. It supports mainstream inference frameworks and development environments including vLLM, SGLang and PyTorch, and provides an OpenAI API-compatible service interface. The accompanying toolchain covers quantization, conversion, compilation, profiling and debugging, lowering the barrier to model migration and application deployment for out-of-the-box use.

From the Cloud to the Site: Bringing Intelligence Back to Where Things Happen

Axera's view is that the value of edge AI lies not in chasing peak numbers on paper but in producing reliable results in real environments. Whether it is spotting surface defects on steel plates, scheduling bins and production lines in a warehouse, or companionship, rehabilitation and care in human-robot collaboration, AI has to sense, reason and respond within the window in which the event occurs.

AI computing foundations like the AX8860 series are becoming infrastructure for edge-side scenarios, supporting the push of model capabilities closer to the ground: computing happens where the data is generated, judgments are made on site without a round trip to a remote server, and instructions reach equipment directly, allowing AI to deliver value in the last mile. For smart manufacturing, that compresses the chain from seeing a problem to assessing it to acting on it, letting on-site systems move from passive response toward a more real-time, more coordinated and more autonomous production model.

Looking ahead, Axera says it will keep building out its full-scenario computing supply and easy-to-use toolchains, and work with industry partners on real business problems, combining AI capabilities with sector know-how, business innovation and delivery services so that underlying computing power better serves industrial upgrading.

As Qiu put it: "Computing power doesn't have to be far away. Intelligence should be right in front of you."