AI tools for Benchmark

Hyra Results is Tencent Hunyuan's repository of released scientific models and reproducible research artifacts. Its model directories provide usable packages and checkpoints for

Xiaomi-Robotics-1 is a vision-language-action foundation model for mobile manipulation. Xiaomi trained the model on more than 100,000 hours of real-world manipulation trajectori

WorldSculpt reconstructs complex scenes as collections of complete object meshes that can be edited separately. It targets densely cluttered environments containing hundreds of

RSI-Exam evaluates whether language-model agents can improve an executable method through sustained experimentation. Instead of asking for a one-time answer or changing the unde

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

SkySynth, also called SkyDiscover-Synthesize, uses coding agents to build systems specialized for a known workload and deployment environment. It targets situations where a gene

Real-SWE evaluates coding agents on engineering tasks drawn from private production codebases. Its tasks involve business logic, existing architectural conventions, and operatio

LLM-as-a-Verifier is a general-purpose verification framework that gives fine-grained feedback for agent outputs across modalities. It treats verification as a scaling axis for

Mage-Flow is a Microsoft research model family for native-resolution image generation and instruction-based image editing. It is a compact 4B model family designed to produce hi

Laguna S 2.1 is Poolside's coding and reasoning model release focused on longer-horizon software engineering work. It is positioned as a significant improvement for models that

Motion4Motion is a training-free motion transfer framework for animating a target subject with motion extracted from a source video. Instead of relying on a shared skeleton, it

Lucida is an open image matting and background removal model focused on hard cases that general-purpose removers often miss. It is tuned for semi-transparent objects, camouflage

Bonsai 27B is PrismML's low-bit multimodal language model family based on Qwen3.6 27B and designed to run locally on consumer hardware, including phones. The release includes te

Inkling is Thinking Machines Lab's open-weights multimodal foundation model for customization, fine-tuning, and agentic workflows. The release describes a Mixture-of-Experts tra

Agents-A1 is a 35B mixture-of-experts agentic language model designed for long-horizon search, engineering, scientific research, instruction following, and tool calling. It targ

OmniContact is a contact-flow framework for generalizable humanoid loco-manipulation. It lets a robot combine reusable meta-skills such as carrying, pushing, sliding, relocating

Seed2.1 is ByteDance Seed's next-generation model family for real-world productivity, offered in Pro and Turbo sizes. It targets general agents, code engineering, office work, r

QUEST is an open family of 2B to 35B deep research agents trained with fully synthetic rubric-tree tasks. The models target fact seeking, citation grounding, and report synthesi

GLM-5.2 is Z.ai's flagship open-weight model for long-horizon tasks, agentic coding, and large-scale engineering work. The release describes a substantial capability jump over G

dots.tts is a 2B-parameter fully continuous, end-to-end autoregressive text-to-speech system. Its backbone combines a semantic encoder, an LLM, and an autoregressive flow-matchi