# Benchmark Environment: Intel Core Ultra 7 258V integrated Arc GPU; OpenVINO GenAI 2026.4.0; single stream; 128 generated tokens; `bench_paged.py`; greedy generation; `ignore_eos=True`. | Build | Runtime | Throughput | | --- | --- | ---: | | LFM2.5-24B-A2B-Uncensored INT8 OpenVINO | GPU | **30.0 tok/s** | | LFM2.5-24B-A2B-Uncensored INT4 OpenVINO | GPU | 37.8–43.6 tok/s | The continuous-batching path is intentionally marked unsupported for this hybrid convolution/state model. The stateful single-stream path is the valid performance measurement.