Honor reveals tiny Mac Studio rival with Nvidia RTX Spark in a 2-liter chassis

Honor has introduced its Tiangong AXB35 family of compact AI workstations, including an Ultra model equipped with Nvidia’s RTX Spark platform. Despite its tiny 2-liter chassis and a weight of around 1.5 kg, the system is designed for large local AI models and multi-agent workloads.

The Tiangong AXB35 Ultra is the most powerful model in the lineup. Nvidia says RTX Spark delivers up to 1 PFLOP of FP4 AI performance, pairing a Blackwell-based GPU with a Grace CPU. The desktop configuration of the chip features up to 6,144 CUDA cores, a 20-core Arm processor and up to 128GB of LPDDR5X unified memory. Nvidia rates the platform at up to 140W, while Honor equips the system with a 2TB NVMe SSD.

Honor positions the Ultra for multi-agent workloads and the local run of large enterprise-class AI models. The company refers to deployment of LLMs with up to 300 billion parameters, although it notes that practical model capacity depends on factors including quantization, context length, concurrency and software.

Honor is planning other versions of the Tiangong AXB35 as well. The Standard model uses an Intel Core Ultra X7 358H with 64GB of memory and a 512GB SSD, while the Pro version is based on AMD Ryzen AI Max+ 395 with 128GB of memory and a 1TB SSD. HONOR lists up to 180 TOPS for the Intel configuration and 126 TOPS for the AMD model, which should be enough for agent workloads and smaller LLMs.

Each version uses the same compact 70 × 192 × 203 mm chassis, with operating modes rated at 55W, 85W and 132W, plus an external 240W power supply. Connectivity includes two 10GbE ports, Wi‑Fi 7, two USB4 ports, four USB-A ports, HDMI 2.1 and DisplayPort 1.4. Storage can also be expanded through three M.2 slots.

The AXB35 supports both Windows 11 and Ubuntu, making it suitable for a range of AI and workstation environments. Honor has not announced pricing or retail availability, and potential customers are currently directed to contact the company about configurations—suggesting the systems may initially be aimed more at the enterprise market.