DF35 is a 4U rack-mount appliance with support for Qwen3.6-35B-A3B MoE and Qwen3.6-27B Dense models. PerfXLab’s PerfXLM inference framework optimizes the hardware and software together.
Performance breakthroughs: synergetic optimization of software and hardware
Best cost and performance
DF35 uses a 4U standard frame design that integrates the Qwen3 series of mainstream large models and achieves depth synergy optimization of software and hardware through a self-researched PerfXLM inference framework. The structure achieves the optimal balance between performance and cost by keeping the per capita deployment cost of a 50-person enterprise at approximately $600, while ensuring the accuracy of the FP16.
It's stable.
In continuous load testing, the equipment showed excellent stability. In the case of the Qwen3-Next-80B-A3B model, the output token rate (TPS) remained above 85 per cent of the single simultaneous processing in the 10 side request environment.
More appropriate prices
A breakthrough pricing of $29,000, bringing the privatization of the Qwen3-Next 80B FP16 precision model into the “Twenty Million Era”, provided a professional and reliable AI infrastructure solution for SMEs.
Planning for the future: ongoing technological evolution
It's so quiet and far away.
A static tower version dedicated to the optimization of the office environment will be launched in the fourth quarter of 2025, while adaptation of larger-scale models, such as Question 235B, will be initiated
Strategic outlook for 2026
In 2026 it is planned to introduce more performance and lower prices and to continue deepening solutions in vertical areas such as manufacturing, finance and health care