September 18, 2026
Running AI locally fits in 2 liters: HONOR AXB35 uses NVIDIA RTX Spark
On September 18, 2026, HONOR unveiled the AXB35 Ultra with 128 GB of unified memory in a chassis with a volume of about 2 L. The workstation uses NVIDIA RTX Spark and claims up to 1 PFLOP at FP4 for running and developing AI systems locally. HONOR also states support for models with up to 300B parameters, but results depend on quantization, context length, parallel workloads, and the framework.
AXB35 comes in three versions: Standard with an Intel Core Ultra X7 358H, 64 GB of memory, and a 512 GB SSD; Pro with an AMD Ryzen AI Max+ 395, 128 GB, and 1 TB; and Ultra with NVIDIA RTX Spark, 128 GB, and 2 TB. For local AI, these are three hardware tiers in one family.
Top model. AXB35 Ultra occupies about 2 L, weighs about 1,5 kg, and measures 70×192×203 mm. The NVIDIA RTX Spark platform claims performance of up to 1 PFLOP at FP4 and support for models with up to 300B parameters.
Connectivity includes two 10GbE ports, Wi-Fi 7, two USB4 ports, four USB-A ports, HDMI 2.1, and DisplayPort 1.4. AXB35 runs Windows 11 and Ubuntu, uses two M.2 2280 PCIe 4.0 slots and one M.2 2242/2280 PCIe 4.0 slot, and offers 55, 85, and 132 W power modes with a 240 W external power supply.
Testing. HONOR claims its own Model Infra and accelerated inference with CUDA, Vulkan, ROCm, OpenVINO, ONNX Runtime, and llama.cpp. In HONOR's own test, the AMD Ryzen AI Max+ 395 achieved 442,66 tokens/s Prefill and 39,88 tokens/s Decode for Qwen3.5-122B-A10B. Compared with an unnamed system, the figures are 24,7% and 65,7% higher.
The price and sales launch date for AXB35 had not been announced as of September 18, 2026.
The next milestone for AXB35: the announcement of its price and sales launch date.
Source
