We generalize the results about how to compute iterative decoding thresholds over the binary erasure channel of non-binary LDPC code ensembles of [16] to non-binary TLDPC codes [2], [3]. We show in this case how density evolution can be performed in order to calculate iterative decoding thresholds and find several famillies with a very simple regular structure and thresholds close to the Shannon limit. To check the performances of these codes over other channels we have tested one of the simplest codes over F-4 which has rate 1/2 on the Gaussian channel. For the (binary) length 1008 for instance, without any optimization on the permutation structure of the code, it matches the performances of the best binary codes of the same length up to the word-error rate 10(-3). We also notice that all LDPC codes (binary or not) having at least two symbols of degree 2 per parity-check equation can be represented as a special kind of TLDPC codes. We show that this representation and the associated decoding algorithm leads in the case of cycle codes to a significant reduction of the number of iterations which are needed for iterative decoding.