---

license: apache-2.0

base_model: nabin2004/AOS-gemma4-manim-merged

library_name: gguf

pipeline_tag: text-generation

language:

- en

tags:

- manim

- gemma4

- gguf

- ollama

- llama.cpp

- tool-use

- animation

- sft

---

AOS Gemma 4 Manim SFT (GGUF)

Q4_K_M GGUF of the AOS Manim SFT merged model. Drop-in for Ollama and llama.cpp.

Model URL: https://huggingface.co/nabin2004/AOS-gemma4-manim-gguf

Files

| File | Description |

|------|-------------|

| aos-gemma4-manim-Q4_K_M.gguf | Quantized weights (~2 GB for E2B) |

| Modelfile | Ollama import template |

Related repos

| Artifact | Repo |

|----------|------|

| LoRA adapter | nabin2004/AOS-gemma4-manim-sft |

| Merged HF weights | nabin2004/AOS-gemma4-manim-merged |

Ollama

Pull from Hugging Face or create locally from the downloaded GGUF:

ollama create aos-gemma4-manim -f Modelfile
ollama run aos-gemma4-manim

Requires Ollama 0.30+ (native Gemma 4 support).

OpenAI-compatible API

Ollama exposes /v1/chat/completions on port 11434:

from gemma4_client import DEFAULT_OLLAMA_BASE_URL, Gemma4Client

client = Gemma4Client(
    model="aos-gemma4-manim",
    base_url=DEFAULT_OLLAMA_BASE_URL,
    api_key="ollama",
)
print(client.chat("Animate a unit circle morphing into an ellipse."))

See apps/server/README.md.

llama.cpp

./llama-server -m aos-gemma4-manim-Q4_K_M.gguf --chat-template gemma --port 8080

How this was produced

cd apps/sft
export LLAMA_CPP_DIR=~/llama.cpp
uv run python export_gguf.py \
  --model-dir ./gemma4-manim-merged \
  --output-dir ./gemma4-manim-gguf \
  --push-to-hub

Quantization chain: merge_adapter.py → convert_hf_to_gguf.py (F16) → llama-quantize Q4_K_M.