AI tools for Evaluation

Ripple-65B is Ricursive Intelligence's open-weight language model for general-purpose reasoning and agentic coding. It is a dense 65-billion-parameter transformer derived from t

STRATA is NVIDIA's global atmospheric emulator for research on storm-resolving simulation. It learns to predict the evolving three-dimensional atmospheric state represented by t

ReLift reconstructs dense 3D point trajectories from a monocular video. It combines the output of a 2D tracker with a video geometry estimator to recover how points move in spac

Phonic-1 is a hosted speech-to-speech model for voice agents. It receives the caller's audio directly and responds in spoken audio, using cues such as pace and hesitation to sup

ToolSpace is a research toolkit for turning tool-use traces into executable training and evaluation environments. It focuses on long-horizon agent tasks that require several sta

Inkling 2 Small is Thinking Machines Lab's multimodal language model for general assistance, coding and tool-using applications. It accepts text and images and generates text, w

Hydrogen-1.0 is Deep Principle's language model for scientific problem solving. It is built on GLM-5.3 and targets tasks such as explaining derivations, organizing research ques

CoLa3D reconstructs an object-composed 3D scene from a single image. It outputs individual meshes that share a coordinate frame so the resulting objects can be inspected and man

Audio8-TTS-Realistic-Air is a compact speech-generation model with multilingual synthesis and zero-shot voice cloning. Its 0.6-billion-parameter main model supports 23 languages

Nanbeige4.3-33B is an open-weight language model for coding and professional agent workflows. It is the flagship teacher model in the Nanbeige4.3 series, combining task-oriented

Groot is an open-weight model family for reasoning and tool-using workloads. The release offers several size and precision variants, including the compact Groot-19B-A3B model, s

Gemini 4 Carbon is Google DeepMind's frontier model for sustained engineering, reasoning, and multimodal work. The initial rollout focuses on selected cyber defenders and truste

Hi3D 1.0 is a publicly released image-to-3D model, formerly called Sparc3D, for turning reference imagery into detailed 3D assets. The model builds on training with nearly two m

FastH3 Omni is an open distilled MiniMax H3 model for reference-guided video with synchronized audio. FastVideo reduces the number of transformer evaluations and provides deploy

MemT-1.0-Nano is a memory foundation model designed to retain useful history as persistent model state. It combines dynamic local memory and trainable hyper memory so historical

Unpaired Rosetta is a cross-modal embedding-alignment implementation that does not require paired examples. Wasserstein Procrustes alignment estimates an orthogonal map between

AlphaProtein Novo is a generative modeling pipeline for de novo enzyme design around structural motifs. A diffusion model jointly generates protein structures and sequences cond

Claude Opus 5.5 is Anthropic's model for complex coding, knowledge work, and sustained professional tasks. It extends the Claude 5.5 family with improved engineering performance

Limite 1B Violetto is Paradigma's small language model for high-throughput mathematical problem solving. A one-billion-parameter dense autoregressive transformer combines curate

ORBIT is a benchmark for evaluating camera-pose estimation in difficult real-world video. It uses panoramic footage to create perspective-view test clips with recoverable refere