Three audiences · one mission

AI that uses less electricity - and proves it.

9.15% less board energy · 91.78% of vLLM batch-invariant throughput†

3.10% less board energy · 80.60% of default vLLM throughput†

Faithful full-model Qwen2-7B CUDA execution now running on NVIDIA H100, with 1.0 token agreement in the current Hugging Face acceptance test.

Third-party tested on H100. Lower board energy per prefill than vLLM. Faithful Qwen2-7B CUDA correctness gates passed.

†TESTfort third-party measurement (Version 99, 2026-07-23). Qwen2-7B-Instruct packed prefill, batch 16, sequence length 128, one NVIDIA H100 80GB. Work unit: prefill positions - not decode tokens or full serving throughput. LuxiEdge medians: 28,374.7 positions/sec, 0.018718 board J/position. Default vLLM medians: 35,203.1 positions/sec, 0.019316 board J/position. LuxiEdge achieved 80.60% of default vLLM throughput and used 3.10% less board energy per position. This is not a throughput win. Formal signed narrative pending; raw artifacts and technician attestation available. Board energy via NVML - not facility or wall-plug energy. Results apply to the tested configuration and workload.

**Current acceptance testing includes matched CPU, CUDA, and Hugging Face greedy-token comparisons under the defined limited protocol. It is a correctness result, not a broad model-quality or performance ranking.

Full methods and downloads: Proof.

Measurement scope: one NVIDIA H100 80GB · Qwen2-7B-Instruct · batch 16 · sequence 128 packed prefill positions · NVML GPU-board energy. Not decode tokens, full serving throughput, facility power, or wall-plug energy. Results apply to the tested configuration. Detailed methods →