# Genie 4, LFM3, Google's sleep framework, Ai2 novelty judges > A new DeepMind world model, open 2B-class models from Liquid AI, a continual learning idea from Google Research, and a study of LLM novelty judges. Published 28 Oct 2026. Coverage: 27 Oct 2026 to 28 Oct 2026. By Ali Asif. ## Items 1. **Genie 4: A new generation of world simulation** (Google DeepMind, Image, video, world): Natively multimodal world model that generates audio and dialogue with the world and can be steered by text in real time, in research preview for trusted testers with a waitlist. [Primary source](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation) 2. **Introducing LFM3: Every Modality for Specialized Intelligence** (Liquid AI, Open weights): Six open-weight 2B-class models on one backbone (text, vision, audio, omni, decision), with the 1.7B text model fitting in about 1 GB at 4-bit and a 256K context window. [Primary source](https://www.liquid.ai/blog/lfm3) 3. **The sleep framework: A new way for LLMs to retain knowledge in continual learning** (Google Research, Science): Proposes alternating a wake phase with an offline sleep phase that consolidates context into weights, reporting retention of long-context knowledge up to 10M tokens. [Primary source](https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning) 4. **Can we trust LLMs to judge whether a research idea is novel?** (Ai2, Science): Benchmark, data and code for novelty judges; a few words changed in a prompt moved accuracy by more than 50 percentage points across six LLM judges. [Primary source](https://huggingface.co/blog/allenai/novelty-bench) 5. **Memory-Enforced GPU Sharing in Kubernetes with NVIDIA CUDA 13.4** (NVIDIA, Agents and coding): CUDA 13.4 enforces per-workload GPU memory limits, and an open-source GPU Fractioning Operator plus KAI Scheduler v0.18.0 expose it on Kubernetes. [Primary source](https://developer.nvidia.com/blog/memory-enforced-gpu-sharing-in-kubernetes-with-nvidia-cuda-13-4) Five primary pages today: a world model, a family of small open models, a continual learning proposal, a benchmark study, and a GPU scheduling release. I opened each one and ran none of them. ## Genie 4, a world model with sound Google DeepMind announced Genie 4, its new world model. The pitch is realism and real-time interaction. It is natively multimodal, so music, ambient noise and dialogue are generated together with the world rather than added afterwards. DeepMind says it builds on the Gemini family: Gemini Omni anchors the simulation in real-world appearance and behavior, and other Gemini models supply the reasoning that keeps the scene coherent. You steer it with keyboard controls for movement and camera, and with text prompts that change the weather, add a person or object, or move the view. Gemini adapts the prompt to the surrounding scene and keeps the world progressing between inputs. DeepMind names training environments for robots and agents as one use, alongside interactive experiences. Access is limited. It is a research preview for trusted testers, with a waitlist in Project Genie, and DeepMind says consumer and developer access comes later, starting with Google AI Ultra subscribers and early access partners of the Gemini API and Google AI Studio. I have no hands-on view yet. ## LFM3, small models for every modality Liquid AI released LFM3, a family of six open-weight models in the 2B range, all branched from LFM3-1.7B-Base: text, vision, audio, omni, and two decision models. The base and post-trained weights are on Hugging Face. Liquid says the 1.7B model outperforms its earlier LFM2.5-2.6B with 35 percent fewer parameters, fits in about 1 GB at 4-bit precision, and has a 256K context window. The architecture mixes gated short convolutions, sliding-window attention and global attention. Of 24 layers, 13 are convolutions, 5 are sliding-window with a 256-token window, and 6 are global attention. Liquid ran its hardware-in-the-loop architecture search across 14 device and backend combinations, including phones, a Raspberry Pi 5, laptops and a Qualcomm NPU. Pre-training used 22T tokens. At 8K tokens of context it reports about 25 percent faster inference than LFM2-1.2B at matched per-parameter performance. ## Google's sleep framework Google Research proposes a sleep phase for continual learning. In the wake phase a model handles inputs and keeps short-term information in fast memory. In the sleep phase it stops taking input, opens up pre-allocated Mixture-of-Experts memory blocks and updates its weights, using consolidation by distillation and a dreaming loop of self-generated practice examples that are filtered and refined. Google says this lets a deployed model absorb new facts and retain long-context knowledge up to 10M tokens while limiting catastrophic forgetting. It is a research framework, not a product. ## Ai2 on novelty judges Ai2 studied how reliably LLMs judge whether a research idea is novel, and released a benchmark, data and code. Across six judges, small prompt changes shifted accuracy on picking the more novel of two ideas by more than 50 percentage points. Retrieval and extra reasoning did not help consistently, and two purpose-built novelty systems did worse than a simple prompted Claude Opus 4.6 baseline. If you use an LLM as a judge, this is a useful caution. ## GPU sharing on Kubernetes NVIDIA's post covers CUDA 13.4, which enforces a memory boundary per workload with a time-sharing mode alongside the existing sm-sharing mode. A new open-source GPU Fractioning Operator and KAI Scheduler v0.18.0 let Kubernetes users request exact memory, or a guaranteed minimum with a burstable maximum. ## What I'd watch next Whether Genie 4 reaches developers and what the API looks like. Whether LFM3 holds up after fine-tuning on real tasks. Whether anyone outside Google reproduces the sleep results. ## My take Genie 4 is the headline, but it is a waitlist preview, so I can only report what DeepMind says. LFM3 is the one I would actually pull down and try on a small machine. I read these pages and ran none of it, so every number belongs to its publisher. ## Sources - https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation - https://www.liquid.ai/blog/lfm3 - https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning - https://huggingface.co/blog/allenai/novelty-bench - https://developer.nvidia.com/blog/memory-enforced-gpu-sharing-in-kubernetes-with-nvidia-cuda-13-4 Canonical: https://heyaliasif.com/ai/2026-10-28-genie-4-lfm3-sleep-framework-gpu-fractioning