Running large language models locally has traditionally demanded massive full-tower desktops or expensive server racks that dominate workspace real estate. Honor is taking a direct swipe at Apple and standard workstation rigs with the Tiangong AXB35, a compact 2-liter desktop machine engineered specifically for local multi-agent processing and massive AI deployments.
Weighing roughly 1.5 kg and measuring just 70 x 192 x 203 mm, the flagship Tiangong AXB35 Ultra packs Nvidia's specialized RTX Spark platform. By pairing a Blackwell-based GPU architecture with a 20-core Arm Grace CPU, the miniature desktop delivers up to 1 PFLOP of FP4 AI compute, opening up high-level enterprise model inferencing to standard desk setups.
Serious AI Muscle in a Pocket-Sized Footprint
The centerpiece of the lineup is the Tiangong AXB35 Ultra, designed to handle AI model parameters that typically choke traditional compact PCs. Inside, the desktop configuration pairs 6,144 CUDA cores with up to 128GB of LPDDR5X unified memory. Honor claims this unified memory pool allows the system to tackle large language models scaling up to 300 billion parameters locally, depending on model quantization, concurrency requirements, and context length limits.
Nvidia rates the RTX Spark compute platform at up to 140W under full load. Honor complements this with an onboard 2TB NVMe SSD and a trio of customizable operating modes rated at 55W, 85W, and 132W. Power is supplied through an external 240W brick, keeping internal thermals focused entirely on cooling the high-density silicon rather than an internal power supply unit.
Flexible Silicon: Intel, AMD, and Nvidia Options
Honor is not limiting the AXB35 line to Nvidia architecture alone. The company is introducing three distinct hardware tiers inside the identical 2-liter chassis to suit varying workload intensities:
• AXB35 Ultra: Nvidia RTX Spark platform (Grace CPU + Blackwell GPU, 6,144 CUDA cores), 128GB LPDDR5X unified memory, 2TB SSD, delivering up to 1 PFLOP FP4 AI compute.
• AXB35 Pro: AMD Ryzen AI Max+ 395 processor, 128GB memory, 1TB SSD, rated for up to 126 NPU TOPS.
• AXB35 Standard: Intel Core Ultra X7 358H processor, 64GB memory, 512GB SSD, delivering up to 180 platform TOPS.
While the Intel and AMD configurations focus on lighter agent workflows and smaller on-device parameter models, they offer standard x86 flexibility for everyday development environments.
Enterprise Connectivity and OS Support
Despite its minimal footprint, the Tiangong AXB35 avoids the single-port compromises common among compact desktop computers. The rear and front I/O panels host dual 10-Gigabit Ethernet ports for high-throughput local network clustering, Wi-Fi 7 wireless connectivity, two USB4 ports, four USB-A ports, HDMI 2.1, and DisplayPort 1.4.
Internal storage expandability is equally generous, offering three full M.2 slots for storage upgrades. The entire lineup supports both Windows 11 and Ubuntu out of the box, allowing developers to immediately run standard Linux AI toolchains or typical enterprise productivity stacks.
Honor has not yet announced official retail pricing or global release schedules for the Tiangong AXB35 series. Prospective enterprise clients and workstation buyers are currently directed to contact the company directly for commercial availability.
