Lafa, an 8B-parameter model, speeds up text generation by 2.54x without sacrificing quality

Continuous flow language models are opening a fundamentally new path toward parallel text generation. Unlike traditional autoregressive models, which generate tokens sequentially, flows offer control and acceleration tools that discrete diffusion approaches can't match. Until now, though, efficient methods for building such models at the scale of billions of parameters have remained underexplored.

Researchers have introduced Lafa (Large flow language model) — the first fast, high-performance flow-based model scaled to 8 billion parameters in a data-efficient way. The work shows that flow-based generation can be a practical alternative to autoregressive modeling at industrial scale.

Architecture and design

Lafa uses an innovative blockwise flow architecture. Causal attention is applied across the prefix (as in standard transformers), while bidirectional asynchronous flow operates inside each block. This hybrid approach achieves two critical goals at once: efficient training by adapting pretrained autoregressive models, and fast inference thanks to prefix caching and blockwise parallel decoding.

Adapting existing AR models means the researchers didn't need to train Lafa from scratch — they could reuse already-trained weights, substantially cutting computational costs.

Training and scaling

The authors trained Lafa in three scaled variants: • 0.5B parameters — a lightweight version for resource-constrained devices • 7B parameters — a mid-scale option balancing performance and speed • 8B parameters — the full-featured version for maximum quality

All three variants were trained on fewer than 30 billion tokens. By comparison, typical diffusion models require significantly more data. This efficiency addresses one of the key challenges of flow-based models — high training costs.

Quality evaluation and practical tasks

Lafa was systematically evaluated across three task categories that determine a language model's practical usefulness: • Instruction following — how accurately the model executes user instructions • Reasoning — the ability to perform logical inference and solve math problems • Coding — generating, analyzing, and debugging source code

On all of these tasks, Lafa's flow-based generation matches or exceeds the quality of traditional autoregressive models. The model can also be distilled into flow maps — compact representations for one-step and multi-step generation, opening the door to even faster variants.

Self-speculative decoding: the key to the speedup

A critical finding from the authors: flow training not only doesn't destroy the base model's autoregressive capabilities, it actually improves them through the multi-token prediction effect. The researchers translated this insight into a self-speculative decoding scheme: flow drafts are verified autoregressively inside a single model, with no switching between separate models required.

The result is impressive: Lafa achieves a 2.54x throughput speedup over a standard autoregressive model at comparable quality. That outperforms both specialized multi-token prediction methods and discrete diffusion approaches.

What this means for language modeling

These results demonstrate that continuous flows are ready for industrial deployment. Flow-based models are no longer just an experimental approach — they can serve as a full-fledged, scalable foundation for the language modeling of the future, combining the speed of parallel generation with the quality of autoregressive approaches.