We present DFM-TTS, a non-autoregressive framework that applies discrete flow matching directly to multi-level codec tokens and augments training with two lightweight guidance signals. First, a text-guided consistency head (auxiliary CTC) tightens semantic–acoustic alignment. Second, a coarse-to-fine, time-dependent weighting stabilizes multi-codebook learning without modifying the sampler. Under zero-shot evaluation on LibriTTS, DFM-TTS improves intelligibility and speaker similarity over a retrained F5-TTS baseline while maintaining competitive naturalness, especially in the few-step regime (NFE ∈ {4, 8, 16}). Because the guidance acts only during training, per-step inference cost remains unchanged. Overall, the approach offers a simple, parallel path to step-efficient zero-shot TTS in discrete token space.