AI Today: Google Genie 4 and Gemini 4 Argon

State of knowledge as of 2026-10-28. Source material collected on 2026-10-28. The day brought a series of technology premieres, including the debut of the Genie 4 model for interactive simulations and the advanced Gemini 4 Argon from Google. Concurrently, AMD introduced support for the Ray library in ROCm 10.0.0, Liquid AI released the versatile LFM3 model family, and the Anima platform implemented OpenAI solutions to optimize work in the medical sector, significantly reducing patient service time and clinical documentation processing.

Highlights Today (TL;DR)

• Google Genie 4: A new multimodal model enabling the generation of realistic, interactive simulations that respond to natural language.
• Gemini 4 Argon: A model designed for complex multi-step tasks, including 3D visualization generation and large dataset analysis.
• AMD ROCm 10.0.0: A software update introducing support for the Ray 2.58.0 library for Instinct series accelerators.
• Liquid AI LFM3: A family of compact models with a size of 2B parameters, optimized for performance on edge devices.
• Anima: The deployment of GPT-6 and GPT-Live models in medical facilities, automating patient interviews and document workflows.

Google Presents Genie 4

Google announced on October 27 the introduction of Genie 4, a new world model designed to generate high-realism interactive simulations. The system is natively multimodal, meaning that in addition to visuals, it generates ambient sounds, music, and real-time interactive dialogues. The model is based on the Gemini family architecture. The Gemini Omni component is responsible for the realism of appearance and behavior within the simulation, while other models from this family provide the reasoning necessary to maintain the consistency and continuity of the world. Genie 4 allows users to initiate events to which the model responds with environmental reactions, moving characters, and sound. During the simulation, it is possible to modify the scene using natural language, for example by changing the weather, camera position, or adding new objects. The user can navigate in the generated environment using standard keyboard controls. Currently, Genie 4 is being made available as a preview version for selected testers. Access for a broader group of users begins with Google AI Ultra subscribers as part of Project Genie and partners enjoying early access to the Gemini API and Google AI Studio. Interested parties can sign up for the waitlist via the project website. For AI Builders: Teams working on robot training environments and creators of interactive experiences can check out the tool in early access. The announcement does not publicly specify the costs of using the model after the testing phase, limitations on the length of generated sequences, or the availability of full functionality for users outside the Project Genie program – this information requires verification in the API documentation.

Google Presents the Gemini 4 Argon Model

Google released a new model, Gemini 4 Argon, on October 27, 2026, designed for complex multi-step tasks and multimedia handling. The tool was created to solve advanced research and programming problems, enabling tasks such as generating 3D visualizations, creating interactive animations, and developing applications based on simple prompts. The model utilizes advanced reasoning and visual perception for long-term workflows. In internal testing, Gemini 4 Argon generated 5,000 lines of mathematical code to render a 3D scene without using external libraries. Other applications include transforming data, such as 80,000 files from Mars missions, into management reports, or creating physical simulations, including engines for rendering hair dynamics. Gemini 4 Argon is available to developers and enterprises via the API in AI Studio and the Gemini platform. The model can be used by Google AI Ultra and Pro subscribers within the Gemini app, the Google Antigravity platform, and the AI Mode feature in the search engine. For AI Builders: Developers working on multimedia applications can test the model's capabilities to generate 3D rendering code and interactive physics elements. The press release does not specify the model's availability in Polish, request limits within individual subscription plans, or detailed pricing for enterprise solutions, which should be verified before starting integration.

AMD ROCm 10.0.0 Supports Ray 2.58.0

AMD announced on October 27 the introduction of support for the Ray 2.58.0 library within the ROCm 10.0.0 software package, enabling the scaling of artificial intelligence workloads on AMD Instinct accelerators. The new version allows existing Ray code to run on Instinct chips without requiring changes to the device code. This ensures support for processes including data loading, distributed training, and serving Large Language Models (LLMs). The validated technology stack includes PyTorch 2.12, vLLM for ROCm, and support for current Instinct GPU models, including the MI355X unit. Thanks to native support for specific accelerator types, such as MI300X, MI325X, and MI355X, systems can recognize specific GPUs in clusters. This translates into the ability to manage heterogeneous device fleets, where individual tasks are routed to the hardware they were tuned for. The publication presented an example of running an OpenAI API-compatible LLM endpoint using vLLM and Ray Serve, where tensor parallelism splits the model across devices, and communication is handled by the RCCL library. Kernel and driver optimizations in ROCm 10.0.0 aim to reduce operational overhead and accelerate data transfer between the host and the device. The Ray library is an open-source framework used to scale Python applications across CPU and GPU clusters, utilized for hyperparameter tuning and reinforcement learning, among other things. For AI Builders: Engineering teams managing distributed GPU clusters can test migrating existing Ray code to the Instinct architecture by configuring task scheduling. The material does not specify third-party software license requirements, long-term technical support for older hardware versions, or infrastructure maintenance costs in various cloud environments – these issues require verification before proceeding with integration.

Liquid AI Introduces the LFM3 Model Family

Liquid AI announced on October 27 the premiere of the new generation of LFM3 models, which integrate text, vision, audio, and decision-making functions within a single architecture with a size of about 2B parameters. These models utilize loop hardware architecture search, allowing them to run efficiently on diverse hardware, from phones and Raspberry Pi 5 to laptops and Qualcomm Hexagon NPUs. The foundation of the series is the LFM3-1.7B-Base model, featuring 24 layers and trained on 22T tokens. Compared to the previous generation, the architecture has been redesigned for agentic tasks requiring fast inference with a long context, which in this model reaches 256K tokens. The manufacturer declares that when handling 8K tokens, the model is about 25% faster than its predecessor while maintaining similar performance. The family also includes:

• LFM3-VL-2B: A vision model with a size of 1.9B parameters, which is the company's first to support video, object recognition, and step-by-step reasoning. The model's results across 62 vision benchmarks achieved an average of 71.4, which, according to the creators, places it on par with the Gemma 4 12B model.
• LFM3-Audio-2B: A model supporting speech recognition, synthesis, translation in 18 languages, and audio reasoning. It utilizes a new audio encoder and codec, achieving a UTMOSv2 score of 3.83 in speech quality. The models are available via the transformers library. In agentic tests such as WebArena-Verified, the LFM3-1.7B model scored 38.42, outperforming LFM2.5-2.6B and Qwen3.5-9B in long-distance tasks performed in a browser and business applications. For AI Builders: Teams deploying local agentic systems can test these models regarding tool support and function calling on their own hardware. The announcement does not provide information on the commercial license, the full list of supported languages, and infrastructure maintenance costs when using proprietary data – these details must be verified before launching production.

Anima Uses OpenAI Models to Automate Patient Care

The Anima medical platform has deployed GPT-6 and GPT-Live models to automate the processing of clinical documentation and handle phone calls in medical facilities. According to information from October 23, the system currently serves over 14 million patients previously handled by more than 1,000 clinics, mainly in the UK, with plans for expansion into the US and Australian markets. At the heart of the solution is the voice assistant Andi, powered by GPT-Live, which conducts medical interviews. The conversation with the patient is guided by a clinical policy engine, and the GPT-5.4 mini model selects the appropriate questionnaires. The collected data goes to the doctor's dashboard in a structured format. Thanks to this solution, clinics have reduced the waiting time for a connection from 40-50 minutes to 10 minutes, representing a reduction of at least 75%. In satisfaction surveys, 86% of patients declared their willingness to use the assistant again. In terms of document management, Anima uses GPT-6 to extract key information such as sender details, document type, and event dates. The system automatically matches documents to patients and suggests appropriate clinical codes from a list covering about 800,000 possible SNOMED codes. The use of in-context learning allows the system to adapt to the specifics of each facility without the need for costly fine-tuning of the models. As a result, clinics increased their document processing throughput from 15 to up to 60 items per hour, enabling the rapid clearance of backlogs exceeding 100,000 documents. For AI Builders: IT teams in medical facilities can check how to integrate in-context learning with their own clinical coding systems instead of fine-tuning models. The source does not specify whether the system supports languages other than English, what the total API costs are at this processing scale, and what the issue of compliance with data protection regulations looks like in jurisdictions other than the UK.

Frequently Asked Questions (FAQ)

What capabilities does the Genie 4 model offer?

Genie 4 is a multimodal world model that generates video, audio, and interactive dialogues in real time. Users can navigate the generated environment using standard controllers and modify scene elements, such as weather or camera position, using natural language.

How does Anima use OpenAI models?

The Anima platform uses GPT-6 to extract information from clinical documentation and match medical codes. Additionally, the voice assistant Andi, powered by GPT-Live, conducts medical interviews with patients, which has helped reduce the waiting time for a connection with the facility by at least 75 percent.

On what devices can LFM3 models be run?

Thanks to loop hardware architecture search, LFM3 models are efficient on a variety of hardware. The creators indicate that they can be run on devices ranging from phones and Raspberry Pi 5 to laptops and dedicated processors such as the Qualcomm Hexagon NPU.

Sources

• Genie 4: A new generation of world simulation — 2026-10-27
• 8 examples of Gemini 4 Argon in action — 2026-10-27
• ROCm 10.0.0 Support for Ray 2.58.0: Enabling Scalable AI Workflows — 2026-10-27
• Build specialized multimodal models with LFM3 — 2026-10-27
• Anima helps clinics cut call waits by 75% with GPT-6 and GPT-Live — 2026-10-23 --- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Version before human editing.