Lafa scales continuous flow LMs to 8B, hitting 2.54x AR throughput via self-speculative decoding

Arnav Chavan · hf · 2026-10-21

Researchers from UC Berkeley, Amazon, KAIST and CMU introduce Lafa, a continuous flow language model scaled to 8B parameters. Key points:

• Architecture: blockwise flow design with causal attention on the prefix and bidirectional asynchronous flow within each block, enabling efficient adaptation of pretrained autoregressive (AR) models plus prefix caching and blockwise parallel decoding.
• Efficiency: trained at 0.5B, 7B and 8B scales on fewer than 30B tokens—more data-efficient than prior diffusion LMs.
• Quality: matches or surpasses AR baselines on instruction following, math and coding; flow training preserves or improves pretrained AR capabilities via a multi-token prediction effect, and learned flows can be distilled into flow maps for one- and few-step generation.
• Speed: a self-speculative decoding scheme that verifies flow-based drafts within a single model yields 2.54x the throughput of an AR model at comparable quality, beating multi-token prediction and discrete diffusion.

Code and models are open-sourced (github.com/lafa-gh/large-flow-lms).

Original post →

More from Research

• Ant's PAR trains a 0.5B model to extrapolate state tracking 16x beyond training length — ant-intl · 2026-10-21
• DecisionPivot benchmark: two 27B models 2.6 points apart differ 70.1% vs 44.2% on contrast groups — startlux-models · 2026-10-21
• S-Agent 2.0 uses a general coding agent for spatial reasoning, topping five benchmarks — ropedia-ai · 2026-10-21
• Video-Point Model learns 4D world models via 3D point trajectory forecasting — Byungwoo Jeon · 2026-10-21
• EvaOmni: an evidence-grounded benchmark for audio-visual reasoning in omnimodal LLMs — VisionXLab · 2026-10-21
• DRLMs: parallel draft refinement beats autoregressive and diffusion reasoning baselines — microsoft · 2026-10-21

Research channel →