recursive-oss/rec is an inference engine that runs Recursive's language models locally. The repository's latest release is v0.1.0. The available facts do not spell out exactly what was added or changed in this version. The README describes rec as an engine intended for single-session use. So what can be confirmed here is the engine and the usage the repository documents, not specific release changes.
The engine focuses on running models locally on Apple Silicon and connecting with related tools. According to the README, it ships run configurations for pi and OpenCode, and it also exposes an OpenAI-compatible chat completions API for its own clients. That compatible API is a connection point that accepts OpenAI-format requests. The repository also documents install, run, status-check and update commands, along with contribution guidance. The README states that it was compared against llama.cpp on an M4 Pro using 4-bit weights and several context lengths. Generation throughput excluding prefill is claimed to be 1.4~2.6 times that of llama.cpp. That is the comparison figure presented by the repository.
That said, the throughput lead rests on the README's own measurement description, and the available facts include no independent verification. The measurement conditions are also limited to an M4 Pro, 4-bit weights and several context lengths. The engine's documented scope of use is likewise single-session. Therefore these numbers alone cannot tell you how it performs in other environments or with concurrent sessions. rec provides a runtime tool and integration path for working with Recursive models locally, but the claim of being faster than llama.cpp is best read as the repository's own measurement result until it is independently verified.