Uno: Cerebras and IFM release LLM with lossless acceleration

Institute of Foundation Models and Cerebras Systems introduced Uno (arXiv:2609.04010, September 3, 2026) — an open LLM with two sets of weights: AR weights handle quality, while diffusion weights predict multiple tokens in parallel; according to the authors, the Psi-Spec sampler accelerates inference by up to 3x while preserving the base model's distribution.

Compiled by Sergey KostenchukPublished 2026-09-10Updated 2026-09-18

2026-09-10 Research HuggingFace

🚀 Uno: Cerebras and IFM release LLM with lossless acceleration

Institute of Foundation Models and Cerebras Systems (with co-authors from UIUC, Cornell Tech, Harvard) introduced Uno (arXiv:2609.04010, September 3, 2026) — an LLM with two sets of weights: AR weights handle quality, while diffusion weights predict multiple tokens in parallel. The Psi-Spec sampler accelerates inference by up to 3x without a separate draft model — according to the authors.

🌍 The new class of "diffusion-augmented LLM" eliminates the main drawback of speculative decoding — a separate draft model with its memory and operational complexity. Code and weights are open.

👤 Uno 8B, Uno 1B, and Uno Qwen3 8B weights — on Hugging Face, inference.py — on github.com/ifm-ai/uno: you can measure the speed yourself. According to the authors, 8B Uno outperforms DiffusionGemma 26B and Mercury 2 in tool use, coding, and long-context reasoning.

Source 1: https://s-sahoo.com/uno/ Source 2: https://arxiv.org/abs/2609.04010

Sources