Liquid AI open-sources d1 diffusion LLM, decisions in 8ms with audio-capable d1-omni-600M
thursdai_pod · x · 2026-10-12
ThursdAI interviewed Maxime Labonne, Head of Post-Training at Liquid AI, about the open-source diffusion language model d1.
- Decision models don't emit a single token; they pick from a set, in 8ms on a GPU.
- The audio-capable d1-omni-600M variant was also released.
- Labonne explains why diffusion-style output makes generation cost effectively free.
A useful primer on the non-autoregressive LLM paradigm and its practical tradeoffs.
More from Models
- Meta releases Muse Wonder claiming it beats Astra, observer bets it's benchmaxxed — bindureddy · 2026-10-13
- Mistral Large 4 scores 13.3% on sudo L7, ahead of Grok 4.7 as Claude leads at 45% — echen · 2026-10-13
- ThursdAI recap: OpenAI drops 722 math manuscripts, Haiku 5.5 hits 10 cents — altryne · 2026-10-13
- 800 Blackwells vs 100,000: can labs without GPUs survive the open-weights race? — altryne · 2026-10-13
- Swapping only the harness moves a model 35 points: 8090 Labs tests 7 models across 5 coding agents — pstAsiatech · 2026-10-13
- Koltun's spatial-cognition demo: GPT-4o 'can't find the bathtub'; puzzle training lifts a 9B model from 13% to 83% — le_james94 · 2026-10-13