AI Train2025
A PyTorch DTensor-native SPMD library for training and fine-tuning LLMs, VLMs, diffusion and retrieval models. Integrates with Hugging Face for day-0 model support, provides YAML-driven recipes, DTensor/FSDP2 parallelism and NVIDIA-optimized kernels (Transformer Engine, DeepEP, FlexAttn).
AI Train2025
Enables bidirectional checkpoint conversion between Hugging Face and Megatron formats and provides a PyTorch-native training library with tensor/pipeline parallelism, FP8/BF16 mixed precision, SFT and PEFT (LoRA) support for large and multimodal models.
AI Infra2025
Model-compression toolkit for large LLMs/VLMs that integrates quantization (FP8/INT4/etc.), speculative decoding, token pruning and deployment hooks—designed for end-to-end performance on single/multi-GPU inference workflows and research-to-prod model optimization.
AI Model2025
Performs open-vocabulary promptable segmentation and tracking on images and videos, returning instance masks and unique IDs from short noun-phrase or exemplar prompts. Trained on a 4M-concept, 52M-mask dataset; uses a presence head and a decoupled detector–tracker design, with near real-time inference for multi-object video.
AI Video2026
Provides a DiT-based audio–video foundation model plus an official Python inference and LoRA trainer. Ships multiple production-ready pipelines (text/image/audio→video), checkpoints, and performance optimizations (FP8, distilled pipelines) for high-fidelity synchronized audio–video generation.
AI Dataset2026
Provides 6,000 runnable, operator-level PyTorch tasks for training and evaluating CUDA kernel generation models; each sample includes executable code, operator descriptors, and provenance tags, with execution-driven filtering to ensure reproducibility and contamination control.
AI Train2026
Runs transformer forward+backward training directly on Apple's Neural Engine by reverse-engineering private ANE APIs; includes per-layer ANE kernels, INT8 optimizations and benchmarks. Proof-of-concept only—relies on undocumented APIs, has low utilization and CPU fallbacks.
AI Model2026
An instruction‑tuned Gemma 4 E4B multimodal model on Hugging Face that accepts text, images and audio and generates text; notable for 128K long context support, built-in thinking mode, and an on‑device‑friendly E4B architecture under an Apache‑2.0 license.
AI Model2026
Identifies and surgically removes the internal activation directions that trigger refusal behavior in large language models, with one-click options on a HuggingFace Space or a local Python API. Combines multiple extraction methods (SVD, whitened SVD, sparse autoencoders), reversible steering, and analysis-informed verification to quantify capability and refusal trade-offs.
AI Model2026
Community fine-tuned multimodal Qwen3.5-9B using Claude 4.6 distilled data to change the model's 'thinking' behavior; offers an uncensored 'heretic' flavor with image-text-to-text I/O, benchmark comparisons, and deployment notes for inference frameworks.
AI Model2026
Generate text, images, video, audio and action/robot trajectories from combined text, image, video, audio and action inputs. A Mixture-of-Transformers omnimodal foundation model (Cosmos3‑Nano, 16B params) focused on Physical AI (robotics, AV, simulation) and optimized for NVIDIA GPU runtimes.
AI Model2026
Pretrained image-model checkpoint hosted on Hugging Face by Facebook (Meta) for vision experiments and transfer learning. Includes downloadable weights and metadata under CC BY‑NC 4.0 — suitable for research and prototyping but restricted for commercial use.