Lafa: 8B flow language model hits 2.54x AR throughput at comparable quality
kalyan_kpl · x · 2026-10-21
The paper "Large Flow Language Models" introduces Lafa, a continuous flow language model scaled to 8B parameters. Design: a blockwise flow model with causal attention on the prefix and bidirectional asynchronous flow within each block, enabling efficient adaptation of pretrained autoregressive (AR) models and fast inference via prefix caching and blockwise parallel decoding.
Key results:
- Trained at 0.5B, 7B, and 8B scales with fewer than 30B tokens
- Flow training preserves or improves pretrained AR capabilities via a multi-token prediction effect, leveraged in a self-speculative decoding scheme
- 2.54x the throughput of an AR model at comparable quality, beating multi-token prediction and discrete diffusion
- Matches or surpasses AR quality on instruction following, math, and coding
Authors from UC Berkeley, Amazon, KAIST, and CMU; code and models are open-sourced (github.com/lafa-gh/large-flow-lms, huggingface.co/lafa-hf/models), showing continuous flows can be a scalable foundation for language modeling.
Related event: Lafa scales flow language models to 8B with 2.5x throughput(3 posts)→
More from Research
- Wikipedia editors gut mech interp article, citing preprints and blog posts as unusable — Sauers_ · 2026-10-22
- Yanai Elazar surveys concept removal in activation space at Copenhagen NLP Symposium — yanaiela · 2026-10-22
- HKU's PrismaCoder: Open Multimodal Code Models Rival Proprietary Systems — hkunlp · 2026-10-22
- Microsoft researcher cracks 31-year-old KLS conjecture with AI agents exploring 100+ proof paths — rohanpaul_ai · 2026-10-22
- GDPval benchmark under fire: first seed artifact in public split looks off — vincent_koc · 2026-10-22
- GPT-6-Astra nears saturation on 12 robot benchmarks but burns 9M tokens per task — rohanpaul_ai · 2026-10-22