Reflection 70B, an AI model, has encountered significant performance issues on LiveBench, attributed to problems with its API and HF weights. Despite repeated attempts, the scores claiming it as the best OSS model could not be verified, leading to plans for retraining the model from scratch. The API for Reflection 70B is now available on OpenRouter, and initial tests of the corrected version show impressive results with a recommended system prompt and temperature=0. However, the initial release performed worse than Llama 3.1 70B. The model can run at 200+ TPS, and further improvements in quality are anticipated. Tests on a private API also showed impressive results.