HONOR Tiangong AXB35 Packs RTX Spark and 128GB Memory Into a 2 Liter AI Workstation

HONOR has introduced the Tiangong AXB35 family of compact AI workstations, led by an Ultra model that combines NVIDIA RTX Spark hardware, 128 GB of unified memory, and 2 TB of NVMe storage inside a chassis measuring about 2 liters.

The system is designed for local AI workloads, including large language models, multi agent applications, and enterprise development. HONOR is also preparing Intel and AMD versions of the same platform, giving the AXB35 lineup three different processor options depending on workload and budget.

The AXB35 chassis measures 70 × 192 × 203 mm and weighs about 1.5 kg. Despite its small size, the system supports multiple high speed interfaces and several storage expansion options.

HONOR Tiangong AXB35 specifications

The Ultra model is the most unusual configuration.

RTX Spark combines a Blackwell based GPU with a Grace Arm CPU. NVIDIA’s reference desktop configuration includes up to 6,144 CUDA cores, a 20 core Arm processor, and as much as 128 GB of LPDDR5X unified memory.

HONOR lists the AXB35 Ultra at up to 1 PFLOP of FP4 AI compute, positioning it for much heavier local inference workloads than a typical mini PC.

The company also uses a 300 billion parameter model deployment figure as an example of what the platform can handle. Actual capability will depend on factors such as model quantization, context length, software stack, and the number of simultaneous workloads.

Three power modes are available

HONOR lists operating modes of 55W, 85W, and 132W for the workstation.

The system uses an external 240W power supply, which helps keep the internal chassis compact while providing enough power for demanding AI workloads.

The same enclosure is used across the Intel, AMD, and NVIDIA versions.

Connectivity is extensive for a machine this small. The AXB35 includes two 10 GbE network ports, Wi Fi 7, two USB4 ports, four USB A ports, HDMI 2.1, and DisplayPort 1.4.

Storage expansion is handled through three M.2 slots, giving the system more room for local datasets, model files, and project storage.

Windows 11 and Ubuntu are both supported

HONOR lists support for both Windows 11 and Ubuntu, making the system suitable for developers working across commercial software and Linux based AI environments.

The AMD configuration uses the Ryzen AI Max+ 395, which already targets compact workstations with large shared memory pools. The Intel model uses the Core Ultra X7 358H and comes with a lower 64 GB memory capacity.

The Ultra model is clearly aimed at the most demanding workloads, with its 128 GB unified memory and RTX Spark platform giving it substantially more room for large local models.

HONOR has not announced pricing or a retail launch date for the Tiangong AXB35 family. For now, potential buyers are being directed to contact the company for configuration details.

The compact design shows how quickly local AI hardware is moving into smaller desktop systems. A 2 liter workstation with up to 128 GB of memory and 1 PFLOP of low precision AI performance would have been difficult to package in this size only a few years ago, but platforms from NVIDIA, AMD, and Intel are making these compact systems increasingly practical.