valeriodolci/capy-flash is a C++ engine that runs a 125B Qwen3.8-Flash-Next quantized to NVFP4 on two 16GB consumer GPUs. The latest release is v0.5.4, and the published information contains no changelog for this version. The repository was previously named ninfer-tp2.
MoE is a model structure in which only some of the experts take part in the computation. capy-flash splits the computation across the two GPUs with tensor parallelism and offloads experts to the CPU. The README publishes its own reproduction benchmark comparing against Strata 0.1.44.1 on a setup of two RTX 5070 Ti 16GB cards, a Ryzen 9 9900X, and 96GB DDR5. At a 15k prompt prefill, capy-flash's NVFP4 result is 3,778 tok/s, above Strata NVFP4's 2,703 tok/s, and the time to first token is 1.82 seconds, shorter than 2.42 seconds. Strata's 3-bit IQ3_S, on the other hand, recorded 4,037 tok/s at a 55k prompt prefill and 124.1 tok/s during generation. The authors explain that capy-flash generated 23% more tokens on coding tasks, but the time for 20 runs was 4.28 hours against 4.32 hours. They report using NVIDIA's official NVFP4 expert without re-quantization, and 0 malformed tool calls out of 1,124 tool calls across 20 multi-stage coding tasks.
The comparison is the authors' own reproduction benchmark taken under specific hardware and prompt conditions. The NVFP4 configuration leads on some metrics, but Strata's IQ3_S leads on others, so these numbers alone do not support a claim of an overall performance advantage.