Cosmos-Reason2-8B-GGUF

Cosmos-Reason2-8B-GGUF

NVIDIA Cosmos-Reason2-8B is an 8.767B-parameter open vision-language model post-trained from Qwen3-VL-8B-Instruct under the permissive NVIDIA Open Model License (commercially usable with derivative rights), designed for physical AI, robotics, and embodied reasoning with enhanced spatio-temporal understanding, physics-based common sense, object detection (2D/3D localization, bounding boxes), and 256K token long-context support for video/image inputs at FPS=4. It excels in use cases like video analytics agents (root-cause analysis via NVIDIA VSS Blueprint), data curation/annotation (Cosmos Curator for sensor data), and robot planning (Isaac GR00T-Dreams for synthetic trajectories), generating chain-of-thought responses in <think>/<answer> format for deliberate decision-making in unfamiliar environments across datasets like EgoExo4D, PerceptionTest, IntPhys, and CLEVRER. Optimized for NVIDIA Blackwell/Hopper GPUs (32GB+ VRAM, BF16 on Linux/H100/A100) with Transformers inference (max_tokens=4096), it handles mp4 videos/jpg images for text outputs in robotics, AVs, and industrial ops while noting limitations in complex dynamics like fast motion or occlusions.

Cosmos-Reason2-8B [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):