---

license: other

license_name: nvidia-open-model-license

license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license

base_model: nvidia/Cosmos-Reason2-32B

pipeline_tag: image-text-to-text

tags:

• gguf
• quantized
• llama.cpp
• vision
• reasoning
• physical-ai
• embodied-ai
• robotics
• multimodal
• vlm
• mmproj
• q2_k
• q3_k_m
• q4_k_m
• q5_k_m
• q6_k
• q8_0
• 16gb
• 24gb

base_model_relation: quantized

library_name: gguf

---

Cosmos-Reason2-32B — GGUF

GGUF conversion of nvidia/Cosmos-Reason2-32B,

NVIDIA's open reasoning vision-language model for Physical AI and robotics —

space, time, and physics understanding for embodied reasoning.

Built on Qwen3-VL-32B-Instruct and

post-trained for video/image-grounded reasoning at FPS=4.

> Licensed by NVIDIA Corporation under the NVIDIA Open Model License.

Quants

| Quant | Size | Notes |

|-------|-----:|-------|

| Q2_K | 12.3 GB | Fits a 16 GB GPU — biggest size hit, still useful for reasoning |

| Q3_K_M | 16.0 GB | Sweet spot for 16-20 GB GPUs |

| Q4_K_M | 19.8 GB | Recommended default — fits a 24 GB GPU |

| Q5_K_M | 23.2 GB | Quality bump, needs 32 GB |

| Q6_K | 26.9 GB | Near-lossless |

| Q8_0 | 34.8 GB | Reference quality |

Plus mmproj-Cosmos-Reason2-32B.f16.gguf (1.2 GB) — the Qwen3-VL vision tower

projector required for image and video input. Pair it with any of the quants above.

Quick start (llama.cpp — text only)

./build/bin/llama-cli \
    -m Cosmos-Reason2-32B.Q4_K_M.gguf \
    -p "Explain why a ball rolls down a slope, step by step." \
    -n 200 --temp 0.6

Quick start (llama.cpp — vision)

The mmproj projector enables image and video reasoning. Use llama-mtmd-cli:

./build/bin/llama-mtmd-cli \
    -m Cosmos-Reason2-32B.Q4_K_M.gguf \
    --mmproj mmproj-Cosmos-Reason2-32B.f16.gguf \
    --image path/to/image.jpg \
    -p "Describe what is happening in this scene and predict what comes next."

For video, pass an MP4 with --video. NVIDIA recommends FPS=4.

Quick start (Ollama)

hf download Hob-forge/Cosmos-Reason2-32B-GGUF \
    Cosmos-Reason2-32B.Q4_K_M.gguf Modelfile --local-dir ./model
cd ./model
ollama create cosmos-reason2-32b:Q4_K_M -f Modelfile
ollama run cosmos-reason2-32b:Q4_K_M "Hello"

Note: at time of writing, Ollama vision support for the Qwen3-VL family lags

llama.cpp. For full image/video reasoning, prefer llama-mtmd-cli with the

mmproj file.

Conversion details

• Source: nvidia/Cosmos-Reason2-32B (BF16 safetensors).
• Tool: llama.cpp convert_hf_to_gguf.py

with the Qwen3VLForConditionalGeneration text + mmproj split.

• Chat template (tokenizer_config.json) is embedded in the GGUF — works with the

standard <|im_start|> / <|im_end|> Qwen turn structure.

• Native context length: 262 144 tokens. The Modelfile defaults to a

conservative 8 192 — raise num_ctx if you have the VRAM.

• Recommended sampling for reasoning: temperature ~ 0.6, top_p ~ 0.9.

License

Released under the NVIDIA Open Model License.

You must comply with that license when using these GGUF weights, including the

Trustworthy AI restrictions and the attribution requirement above.

• Hob-forge