Model releases, research, papers, LLMs, reasoning, multimodal, benchmarks
Radical Numerics has developed a model capable of identifying both natural and AI-engineered pathogens without requiring an internet connection. This model represents a significant advancement in biological security, enabling rapid detection of potential threats. It allows relevant authorities to take swift action against any biological risks that may arise. The technology is essential for addressing the growing challenges posed by bioweapons, enhancing nations' abilities to protect their populations from potential attacks. Decision-makers should consider integrating this model into national security strategies.
AlphaGenome 2 has launched a new tool designed to analyze how DNA sequences impact specific human cell types and diseases. This tool aids researchers in predicting how variations in DNA sequences affect cells, enhancing scientific understanding of diseases. This tool represents a significant advancement in precision medicine, potentially contributing to the development of personalized treatments based on an individual's genetic makeup.
Google DeepMind has launched the AI model AlphaGenome 2, aimed at understanding how changes in DNA affect different cells. This model is a successor to the previous AlphaGenome model released last year, providing a more detailed view of a complex problem related to disease causes. AlphaGenome 2 analyzes genetic patterns that vary among different cell types, such as heart muscle cells and skin cells, helping scientists determine how these cells interact with various factors. This advancement represents a significant step in medical research, potentially leading to new treatment discoveries.
Baidu released the ERNIE 5.5 Flash Preview on November 11, marking the first post-trained version based on the ERNIE-5.5-Flash-Base model, which was open-sourced on November 6. The model features 280 billion total parameters, with 14 billion active per token, and supports a context of up to 1 million tokens. Baidu claims that the model excels in programming, 3D modeling, knowledge work, and content creation, achieving competitive benchmark scores with fast decoding. However, the benchmark tables were not available as text in the source material, raising questions about the claims.
MiniMax has announced the release of MiniMax-H3-Image with open weights, a text-to-image and image-editing model. Launched on November 10, the model produces 16-megapixel images in under 10 seconds, which is about five times faster than the market average. The model is based on the H3 design and contains 33.1 billion parameters, accepting up to 16 reference images. It also features a dedicated MiniMax-H3 text encoder and allows output sizes up to 6272 by 2688 without an upscaler.
PSI utilized an unreleased model from Anthropic to address a fundamental question in quantum computing that scientists have pursued for nearly two decades. This initiative is part of PSI's efforts to leverage AI for accelerating scientific discoveries, highlighting the importance of integrating modern technology into research. If PSI successfully develops this model, it could lead to significant advancements in quantum computing, opening new avenues for innovation in science and technology.
NaiveAI has launched its new model, Naive-N0.5-Flash, designed to generate tokens at a speed of 2,100 tokens per second. This model, based on the MiMo-V2.5, aims to enhance engineering and research workflows, allowing teams to experiment and solve problems more efficiently. The model features improvements for long-context reasoning, having been trained on 48 terabytes of data, plus an additional 3.25 terabytes. It utilizes a hybrid design that combines global self-attention with sliding-window attention, enhancing its data processing capabilities.
Nvidia has launched its new model GR00T 2, aimed at teaching robots new tasks without the need for laborious teleoperation. The system transfers human hand movements from videos to robotic arms, trained on approximately 400,000 hours of footage. Success rates vary significantly, ranging from 10 to 80 percent depending on the task. This performance variability indicates that the model still requires improvements to achieve more consistent results.
Google has announced the general availability of the Gemini 4 Carbon model following a two-week testing phase. The model ranks first in several benchmarks, outperforming Anthropic's Fable 5.2 in some cases, while costing less than half per task. Additionally, Google is previewing coordinated agent teams through the '/swarm' feature, which divides complex tasks in real-time. This feature reflects a shift towards enhancing efficiency in handling multiple tasks simultaneously.
Google has announced the public rollout of its latest AI model, Gemini 4 Carbon, after a two-week trial with a select group of "cybersecurity defenders." Carbon is touted as Google's strongest model yet, capable of competing with top systems from Anthropic and OpenAI. Google claims that Carbon "pushes the frontier even further" than its previous model, Argon, which launched just last month.
Anthropic's Fable 5.2 has achieved 41.3% of real freelance jobs at professional quality on the Remote Labor Index, more than double the performance of the next-best systems, Opus 5.5 and GPT 6 Astra. A year ago, the top score was only 2.5%, indicating significant progress in the system's capabilities. However, areas such as product design and animation remain weak spots for the model. This achievement highlights Anthropic's ability to enhance AI models, which could positively impact the freelance job market and promote the adoption of these systems in new fields.
Kora, developed by a cofounder of Cyanogen, is a next-gen operating system designed as an advanced alternative to Windows and macOS for AI users. It aims to meet the needs of users running local AI models, providing an optimal environment for efficient work. As reliance on AI increases, Kora represents a significant step toward enhancing the experience for specialized users in this field.
Noitom Robotics has announced the release of the HiPHI-Hand dataset on Hugging Face, which was used to train NVIDIA's EgoHand model. The announcement was made at the CoRL 2026 conference in Austin, Texas, with the release date set for November 9. The dataset includes 11 hours of dual-fisheye video, capturing 3D hand-joint positions, allowing for high accuracy even when hands are obscured from view.
Anthropic has launched the new version of its model, Fable 5.2, reclaiming the world LLM leadership from Opus 5.5. This release comes with a 20% reduction in token costs but requires more tokens for usage. Reports indicate that Fable 5.2 outperforms previous models, achieving a score of 16.4% in performance tests.
Anthropic has announced the release of Claude Fable 5.2 and Mythos 5.2, its most powerful models to date, available in two versions: one free and one for verified organizations. The new models excel in solving open physics problems, enhancing cybersecurity capabilities, and designing effective operating systems, contributing to over 90% of the changes merged into Anthropic's core code. Anthropic emphasizes that it has deliberately restricted any capabilities that could allow the AI to accelerate its own development, reflecting its commitment to safety and security in AI advancement.
Anthropic has announced the launch of its new models, Mythos and Fable, which show improvements in coding, reasoning, and vision. These models arrive amid growing concerns about repetitive strain injury (RSI) and its impact on users. The company has emphasized that these models enhance clarity in communication with users, facilitating a more effective use of AI. This move is significant in the AI field, as Anthropic aims to improve user experience and increase model efficiency in providing support across various domains, reflecting a continuous push towards innovation and development.
Anthropic has announced the launch of two new models: Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle complex technical challenges. The company claims that these new models show improvements in coding, reasoning, and vision, generating outputs faster than the previous 5.1 versions. Claude Fable 5.2 is particularly effective at solving challenging software engineering problems, orchestrating agent teams, and managing long-term projects. Additionally, the new model has solved complex scientific issues, including two open theoretical physics problems. During prerelease testing, the new models demonstrated promising results.
Following the recent expansion of its Edits app to macOS, Instagram has launched the Effect Builder, enabling users to create custom filters for their Stories using AI. This feature is now available for US residents over 18 and includes a prompt box for users to describe their filter ideas, with an option to attach style reference images. Users can upload their own photos to preview the AI-generated effects and request changes in follow-up prompts. Additionally, there's an option for Instagram's AI to rewrite user prompts, transforming simple requests into detailed outputs.
In approximately one year, Handshake achieved $1 billion in annualized revenue from its AI training business. This analysis explores what frontier labs are purchasing at this scale, focusing on the workings of reinforcement learning models. Demonstrations in finance, life sciences, and coding highlight how models are utilized in complex tasks such as financial data analysis and medical image recognition. The finance demo illustrates how the model can accurately interpret tax transactions.
Disney has announced its collaboration with Adobe to develop custom AI tools for its productions. This partnership aims to create brand-safe AI models that align with Disney's branding standards. Adobe's Firefly is an innovative AI tool that provides advanced content creation solutions. Through this collaboration, Disney seeks to enhance its artistic production using modern technologies that ensure brand safety. This partnership represents a strategic move for Disney in utilizing AI, potentially transforming content production in the entertainment industry. It is expected to improve efficiency and increase creativity in its future projects.
The Walt Disney Company has announced a multiyear partnership with Adobe to develop proprietary generative AI models tailored for Disney's production pipeline. Adobe will provide Disney with early access to its latest technologies, spanning from Disney's theme parks to ESPN content. Disney will collaborate directly with Adobe on generative AI models using Adobe Firefly Foundry, focusing on technologies that are responsibly developed. This partnership aims to explore emerging technologies that can enhance Disney's creative processes.
OpenAI has introduced the Control Room system, which enables tracking of model workloads across its research infrastructure. This system can swiftly halt these workloads when safety or security concerns arise. This innovation marks a significant step towards enhancing the safety of models used in research.
Kuken Industries utilizes the OpenAI API to search through 70 years of cooling tower drawings. Engineers can now locate the drawing needed for repairs in just 4.5 minutes, a significant reduction from nearly an hour previously. This technology greatly accelerates the search process, highlighting the role of AI in enhancing efficiency within engineering industries. Kuken leverages historical data to deliver accurate and rapid results.
Visinex has introduced the VS815V, its first RISC-V edge AI vision chip, at the Beijing security-products exhibition held from November 3 to 6. This chip is positioned for mainstream camera and vision devices, with finished products expected to ship as early as 2027. The chip features a dual RISC-V CPU configuration and an INT8 neural processing unit capable of supporting CNN and Transformer workloads. It specifies four sensor inputs over eight lanes of MIPI or sub-LVDS connectivity, with H.265/H.264 encoding at 4K and 30 frames per second.