leonccaa/revolta is an inference server targeting multiple NVIDIA V100 (SM70) GPUs. According to the v0.1.5 release notes on GitHub, this version fixes DFlash2 so that it touches the entire cache prefix when reusing prompts. It also fixes the index shards that the model loader had been skipping. The release description notes that the accuracy of large integer tool arguments is preserved. These changes concern the prompt reuse path, the model loading path, and the correctness of tool argument handling.
The repository description says it targets Qwen3.8 Flash-Next and lists AWQ and NVFP4 as supported formats. It can accept multimodal input, and it also includes the ability to offload the KV cache to CPU or disk. Speculative decoding is described as supporting the MTP or DFlash2 path. The repository is therefore presented as a server that provides inference for that model in V100 environments together with several memory and decoding paths. The center of v0.1.5 is not the addition of this overall feature list, but the fixes for the reuse cache and index loading problems stated in the release notes.
The provided information contains no throughput or latency measurements and no comparison with other GPUs. No license is presented either, so distribution conditions cannot be judged here. What the release notes make clear is the list of fixes and the preservation of accuracy for large integer tool arguments; the extent of the improvement and the effect in real-world use are not confirmed. It is therefore more accurate to view v0.1.5 as an update that refined cache reuse and model loading within the publicly stated scope of changes.