A post-trained parallel language model reportedly matches autoregressive quality with 2.3–5.7× faster inference

Researchers introduce Parallel Language Models, designed to generate several coordinated tokens in one forward pass instead of one token at a time. Their paper reports that, after post-training, a 1.3B-parameter model matched or exceeded an identically trained autoregressive model on math, coding and instruction-following benchmarks while speeding up inference 2.3–5.7×. Its training objective combines an autoregressive loss with a parallel decoding loss.