We consider phrase based Language Models (LM), which generalize the commonly used word level models. Similar concept on phrase based LMs appears in speech recognition, which is rather specialized and thus less suitable for machine translation (MT). In contrast to the dependency LM, we first introduce the exhaustive phrase-based LMs tailored for MT use. Preliminary experimental results show that our approach outperform word based LMs with the respect to perplexity and translation quality.