AWS has launched the Amazon SageMaker JumpStart Model Leaderboard in SageMaker Studio to help users compare foundation models side by side. The leaderboard reports consistent benchmark scores, performance metrics, and cost estimates across multiple provider models.
AI Model Releases, LLM Updates & Benchmarks
Follow AI model releases, LLM updates, benchmark results, and foundation model research.
Wednesday, November 18, 2026
Google DeepMind has launched Gemini 4 Neon, a frontier-level model optimized for complex software engineering, science workflows, and cyber defense. The model features a 1-million-token context window and strong benchmark performance across multiple domains.
Google officially introduced Gemini 4 Neon, a model designed for complex software engineering, scientific reasoning, and cyber defense applications. It features configurable thinking levels and strong benchmark performance across multiple domains.
Google officially introduced Gemini 4 Neon, a model designed for complex software engineering, scientific reasoning, and cyber defense applications. It features configurable thinking levels and strong benchmark performance across multiple domains.
Google has released Gemini 4 Neon, its fourth new Gemini model in a month, offering frontier-level intelligence and a one million-token context window. The model is available via the API at $3 per million input tokens and $15 per million output tokens.
RadixArk released DaisyRose-27B under the Apache 2.0 license, a model trained via reinforcement learning on long-horizon agent tasks using its open-source Miles framework. The model improves performance on Terminal-Bench and various reasoning benchmarks while lowering training costs.
Google DeepMind has released Genie 4, its advanced world model for generating interactive 3D worlds, available to Google AI Ultra subscribers. The update also brings chainable worlds, longer session limits for Genie 3, and a community gallery for sharing creations.
Researchers analyzed 272 production fine-tuning runs on NVIDIA H200 GPUs to determine why traditional FLOP-based duration estimates fail. They found that a CatBoost regressor incorporating evaluation cadence and checkpoint policies achieved high prediction accuracy, outperforming closed-form math.
Google DeepMind has released Genie 4, its most realistic interactive world model, to Google AI Ultra subscribers within the Project Genie platform. The update features real-time prompt editing, native audio generation, and a new conversational creation agent.
Microsoft has introduced a new lineup of in-house AI models at its Ignite conference, covering coding, cybersecurity, decision routing, image editing, and real-time speech. The release aims to provide cost-efficient alternatives to external partner models while running on Microsoft's custom Maia 200 accelerators.
Microsoft has released a family of seven new models, including MAI Cyber 2 Flash, MAI Code 2, and MAI Image 2.7, focused on high performance and lower costs. Additionally, the company announced that its Maia 200 AI accelerator is now in production.
IBM and Confluent have integrated IBM time-series foundation models natively into Confluent Cloud and Confluent Platform for Apache Flink. The models run on standard CPUs to support real-time forecasting, anomaly detection, and optimization use cases.
AI startup Osmo has developed a foundation model trained on proprietary scent data to simulate human-level olfactory intelligence. The technology aims to help design fragrances, detect threats, and assist in medical diagnostics.
OpenAI released GPT-6.2 Astra and Sol, offering lower inference costs and competitive benchmark scores on coding and agentic tasks. Meanwhile, NVIDIA posted record quarterly revenue, and Isomorphic Labs raised a $5 billion Series C round.
Researchers introduced Energy-Based Reasoners, which learn an energy function over latent reasoning states to formulate looped updates via gradient descent. The models achieve high accuracy on complex algorithmic benchmarks while requiring significantly less inference compute than prior architectures.
OpenAI published an evaluation of how coding agents accelerate AI research and where their performance remains uneven. The findings offer insight into the current capabilities and limits of automated research workflows.
Recent evaluations show that the latest Claude models can accurately predict amino acid sequences for little-studied proteins, outperforming other frontier large language models. This capability indicates that protein sequences formed a sizeable component of their training data at scale.
Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model supporting over 40 languages under the Apache 2.0 license. Built using on-policy distillation, the model achieves competitive performance against significantly larger models while running locally via standard runtimes.
A developer implemented expert streaming for Mixture of Experts models inside the Hugging Face Transformers library to run large models on consumer hardware. The approach streams routed experts from disk through a bounded cache to approach the performance of specialized engines like Strata.
Tuesday, November 17, 2026
OpenAI introduced GPT-6.2 Astra as its new frontier model and GPT-6.2 Sol as an efficient everyday model, both available in the API. Initial community reviews highlight mixed benchmark scores and communication style quirks during complex multi-agent tasks.