Lafa turns pretrained LLMs into flow models: 2.54x faster decoding at 7B scale
KyeGomezB · x · 2026-10-28
A UC Berkeley/Amazon/KAIST/CMU paper introduces Lafa (Large flow language models), converting pretrained autoregressive LLMs into continuous flow models that generate multiple tokens in parallel by progressively denoising embeddings.
- Design: blockwise flow with causal attention on the prefix and bidirectional asynchronous flow within blocks; enables prefix caching and blockwise parallel decoding on top of pretrained AR models.
- Decoding: parallel flow predictions serve as drafts, verified via the same model's original next-token heads (self-speculative decoding).
- Efficiency: scaled to 8B with fewer than 30B training tokens.
- Results: 7B Lafa reaches 2.02–2.54x the throughput of comparable AR models at matched or better quality, outperforming MTP and discrete diffusion; it matches or beats AR SFT (Qwen2.5 7B) and can be distilled for one-/few-step generation.
Code and models are open-sourced.
More from Models
- OpenAI opens Decisions API beta: up to 10x faster decisions than GPT-6 Luna — OpenAIDevs · 2026-10-28
- Open-Weight AI Security Research Stalls as $80 Fine-Tuning Undoes Safety Interventions — Miles_Brundage · 2026-10-28
- Socher's Recursive Unveils Flint: 125+ tok/s Local Coding Model Built by Automated Research — jeffclune · 2026-10-28
- Qwen3.8 Flash on a single DGX Spark gets up to 62.7% faster code generation in new update — EAccelerate_42 · 2026-10-28
- Arena co-founder: US still leads frontier AI, but China is winning open models and driving down cost — JaynitMakwana · 2026-10-28
- Liquid AI launches LFM3: tri-hybrid architecture, six edge models, all modalities — maximelabonne · 2026-10-28