Downloads · 30 days
590
47% of all-time downloads
Cosmos-Reason2-2B-GGUF
hzang/Cosmos-Reason2-2B-GGUF
Cosmos-Reason2-2B-GGUF is a image-text-to-text model from hzang. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for gguf. The card lists the license as other.
llama.cpp-compatible GGUF builds of nvidia/Cosmos-Reason2-2B, a 2.4B vision-language model post-trained from Qwen3-VL-2B-Instruct for physical-AI reasoning, spatial understanding, anomaly detection, and chain-of-thoug…
All-time downloads
1.2K
Public
Repo size
4.3 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.gguf4.3 GB · 100%
At a glance
- Task
- Image-Text-to-Text
- Library
- gguf
- License
- other
- Access
- Public
- Created
- May 2, 2026
- Updated
- May 2, 2026
- SHA
- ff68dd9f
Try a prompt
Base models
Model card
From the Hugging Face model README
Cosmos-Reason2-2B-GGUF
llama.cpp-compatible GGUF builds of nvidia/Cosmos-Reason2-2B, a 2.4B vision-language model post-trained from Qwen3-VL-2B-Instruct for physical-AI reasoning, spatial understanding, anomaly detection, and chain-of-thought scene analysis.
Files
| File | Size | Notes |
|---|---|---|
Cosmos-Reason2-2B-Q4_K_M.gguf | ~1.28 GB | 4-bit k-quant. Recommended for memory-constrained devices (e.g., iPhone 13+, 6 GB RAM). |
Cosmos-Reason2-2B-Q8_0.gguf | ~2.10 GB | 8-bit quant. Recommended where memory allows (iPhone 15 Pro+ / 8 GB+). |
mmproj-Cosmos-Reason2-2B-F16.gguf | ~0.82 GB | Vision projector. Required for multimodal inference; load alongside either quant. |
Usage with llama.cpp
llama-server \
-m Cosmos-Reason2-2B-Q4_K_M.gguf \
--mmproj mmproj-Cosmos-Reason2-2B-F16.gguf \
--host 0.0.0.0 --port 8080 \
-c 4096 --jinja
OpenAI-compatible endpoint at http://localhost:8080/v1. Standard chat/completions with image_url content blocks works out of the box.
Quality
Validated against the BF16 reference (nvidia/Cosmos-Reason2-2B) on a golden eval covering spatial reasoning, scene understanding, counting, OCR, anomaly detection, bounding-box output, and refusal. Q4_K_M scores 6/7 — equivalent to MLX Q4 — with the single discrepancy being an evaluator heuristic artifact, not a real quality regression.
Source / lineage
GGUF builds re-uploaded from Kbenkhaled/Cosmos-Reason2-2B-GGUF for hosting stability under the hzang/ account, alongside the MLX variants:
hzang/Cosmos-Reason2-2B-4bit— MLX 4-bit (1.78 GB)hzang/Cosmos-Reason2-2B-8bit— MLX 8-bit (2.64 GB)hzang/Cosmos-Reason2-2B-GGUF— llama.cpp GGUF (this repo)
License
NVIDIA Open Model License + Apache 2.0.
- Task
- Image-Text-to-Text
- Library
- gguf
- License
- other
- Languages
- en
- Created
- May 2, 2026
- Updated
- May 2, 2026