Models
LLM Report articles in Models, with their sources.
Nov 13
Google releases Gemini 4 Hydrogen, a fast model that replaces 3.8 Flash and upgrades SearchGoogle launched Gemini 4 Hydrogen on 12 November 2026 as its fast, cheaper tier. It reaches the Gemini API, the Gemini app and Search for all users, and replaces Gemini 3.8 Flash across Google's products.
Models4 sources
Nov 11
Google DeepMind releases Gemini Robotics 2.5 to robotics partners with new safety benchmarkGoogle DeepMind says its Gemini Robotics 2.5 model lifts task success rates by 30% to 100% and reached 96% in a Mercedes-Benz factory pilot. It is limited to robotics partners for now.
Models2 sources
Nov 10
Google DeepMind's AlphaProof Code verifies Signal protocol's real implementation with AI agentsGoogle DeepMind says AI agents proved security properties of the code behind the Signal protocol in four weeks, using a new framework called AlphaProof Code. The company is open-sourcing the verification artifacts.
Nov 4
Collinear AI releases CWE-bench v2, a public security-repair benchmark where Gemini 4 Argon leadsCollinear AI opened 100 tasks from CWE-bench v2 to the public on 3 November 2026. In its own tests of 11 models, Gemini 4 Argon scored highest, with Claude Opus 5.5 close behind.
Nov 4
NVIDIA engineer builds talking Halloween garage door with Nemotron open modelsAn NVIDIA employee describes in a Hugging Face post how MonsterHouse, a grumpy pirate-voiced projection, runs on Nemotron 3.5 Bolt E4B 4GB and other NVIDIA models on one workstation, with a guide for building similar characters.
Nov 4
IBM ships Granite Switch 4.1.1 patch with new speech and decision adaptersIBM's Granite Switch 4.1.1 is a backwards-compatible patch that updates a 10-language speech-to-text model and a Decision adapter without retraining the base model, according to a Hugging Face post.