Tag
Explore by tags
A vision-language-action foundation model and reference stack for generalized humanoid and cross-embodiment robot manipulation. Provides pretrained checkpoints, demo datasets, and tooling for fine-tuning, evaluation, and deployment (ONNX/TensorRT); released as Early Access.
Splits LLM inference into separate prefill and decode GPU pools, then routes requests with KV-cache awareness to cut redundant recomputation. Reports up to 30x throughput on DeepSeek-R1 (GB200 NVL72) and works across TensorRT-LLM, vLLM, and SGLang.
A ~5,000-line Python LLM inference engine that re-implements SGLang's serving optimizations — radix KV-cache reuse, chunked prefill, overlap scheduling, tensor parallelism — as a fully type-annotated reference instead of a black box.
Generates real-time, infinite-length portrait video from one reference image on a 12GB GPU. Combines implicit facial signals and 3D keypoints with step-distilled diffusion and autoregressive micro-chunk streaming for low-latency live use.
Enables parallel speculative decoding by using a lightweight block-diffusion draft model to produce multi-token drafts for faster, high-quality generation. Integrates with vLLM, SGLang and Transformers backends and ships draft models on Hugging Face.
Generates high‑fidelity, expressive speech and environmental sounds from text. The MOSS‑TTS Family provides specialized models for long‑form TTS, multi‑speaker dialogue, voice design and realtime streaming, plus torch‑free inference paths (llama.cpp / ONNX) and Hugging Face releases.
Unified multimodal LLM for enterprise workflows: ingests video, audio, image and text to perform transcription, OCR, Q&A, summarization and long-context reasoning. Provides BF16/FP8/NVFP4 weights and integrations with vLLM, TensorRT-LLM and other runtimes.
A 33B Mixture-of-Experts text-to-text model optimized for local, long-context agentic coding—3B activated params per token, 131k token window, mixed sliding-window and global attention, FP8 KV cache, Apache-2.0 license.
Unifies video, audio, image and text understanding for enterprise Q&A, summarization, transcription and document intelligence. The NVFP4 quantized variant reduces footprint to ~20.9GB for more efficient single‑GPU deployment and is tuned for NVIDIA runtimes (vLLM, TensorRT).
Provides a GGUF-quantized build of NVIDIA's Nemotron 3 Nano Omni 30B (Reasoning) for local inference — enables multimodal (video/audio/image/text) reasoning, transcription, and document understanding on compatible runtimes such as llama.cpp, Ollama, vLLM, and TensorRT-LLM.
High-throughput LLM inference engine for agentic workloads, combining a local‑SPMD static compiler for parallelism, a C++ scheduler with a Python execution plane and type‑safe KV‑cache reuse, pluggable high-performance kernels (including an MLA implementation), and a low‑overhead AsyncLLM entrypoint for production GPU inference.
Open-weight frontier LLM for agentic reasoning and long-context analysis (up to 1M tokens). Uses a LatentMoE + Mamba-2 hybrid with Multi-Token Prediction and NVFP4 efficiency (550B total / 55B active). Suited for multilingual agents, RAG, and heavy tool-use workloads.