Lafa: 8B flow language model hits 2.54x AR throughput with under 30B training tokens
LucaAmb · x · 2026-10-24
The paper "Large Flow Language Models" (arXiv:2610.28559) introduces Lafa, a continuous flow language model scaled to 8B parameters, showing continuous flows can serve as a scalable foundation for language modeling.
Design
- Blockwise flow model: causal attention on the prefix, bidirectional asynchronous flow within each block
- Trained by adapting pretrained autoregressive (AR) models; fast inference via prefix caching and blockwise parallel decoding
Results
- Trained at 0.5B / 7B / 8B with fewer than 30B tokens, more data-efficient than prior diffusion models
- Matches or surpasses AR quality on instruction following, reasoning, and coding
- Flow training preserves or improves pretrained AR capabilities via a multi-token prediction effect, enabling self-speculative decoding within a single model
- 2.54x AR throughput at comparable quality, beating multi-token prediction and discrete diffusion
Code and checkpoints (0.5B/7B/8B) are open on GitHub and Hugging Face.
More from Infra
- Ex-Fidelity manager: CoreWeave is a zero short of a bailout, AI is Shale 2.0 — SumitGup · 2026-10-24
- Kimi K3.1 Pro launches with 2.8T params beating Fable 5.1, open weights due Nov 4 — Latent Space · 2026-10-24
- 20VC: Naveen Rao says inference is a $10T market and robotics 100x bigger than coding agents — 20VC · 2026-10-24
- King Slide, the 'king of AI server slide rails,' is up 2,385% in five years — jwt0625 · 2026-10-24
- Semis roar back into VC focus as CASPA panel calls for 1000x faster, greener AI chips — hardimanjames · 2026-10-24
- Dev squeezes a 7M-param byte-level LLM entirely into an RTX 4090's on-chip memory — dejanseo · 2026-10-24