Instructions to use Mathos34400/resilient-challenge-image-to-text with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Mathos34400/resilient-challenge-image-to-text with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Mathos34400/resilient-challenge-image-to-text",
	filename="gemma-4-q4_k_m.gguf",
)

llm.create_chat_completion(
	messages = "\"cats.jpg\""
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use Mathos34400/resilient-challenge-image-to-text with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Use Docker

docker model run hf.co/Mathos34400/resilient-challenge-image-to-text:Q4_K_M

LM Studio
Jan
Ollama
How to use Mathos34400/resilient-challenge-image-to-text with Ollama:
```
ollama run hf.co/Mathos34400/resilient-challenge-image-to-text:Q4_K_M
```

Unsloth Studio new

How to use Mathos34400/resilient-challenge-image-to-text with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Mathos34400/resilient-challenge-image-to-text to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Mathos34400/resilient-challenge-image-to-text to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Mathos34400/resilient-challenge-image-to-text to start chatting

Pi new

How to use Mathos34400/resilient-challenge-image-to-text with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Mathos34400/resilient-challenge-image-to-text:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use Mathos34400/resilient-challenge-image-to-text with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use Mathos34400/resilient-challenge-image-to-text with Docker Model Runner:
```
docker model run hf.co/Mathos34400/resilient-challenge-image-to-text:Q4_K_M
```

Lemonade

How to use Mathos34400/resilient-challenge-image-to-text with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Mathos34400/resilient-challenge-image-to-text:Q4_K_M

Run and chat with the model

lemonade run user.resilient-challenge-image-to-text-Q4_K_M

List all available models

lemonade list

Access to Mathos34400/resilient-challenge-image-to-text

This repository contains a compressed Gemma-4-E4B-it model submitted for the Resilient AI Challenge 2026 (image-to-text category). Access is granted manually. The original Gemma license applies.

By requesting access you confirm that you accept the Gemma license terms (https://ai.google.dev/gemma/terms) and that you will use this model in compliance with it.

Gemma-4-E4B-it — Q4_K_M GGUF + Q8_0 mmproj

Submission for the Resilient AI Challenge 2026 — image-to-text category.

Quantized multimodal version of Google's Gemma-4-E4B-it, packaged for inference on a single NVIDIA L4 (24 GB).

Runtime: llama.cpp (llama-server). Image-to-text inference requires llama.cpp with the included multimodal projector (mmproj). vLLM's GGUF backend does not currently support multimodal Gemma 4 inference, so vLLM cannot be used for image input on this submission.

Files

File	Role
`gemma-4-q4_k_m.gguf`	Language model — Q4_K_M K-quant (4-bit, imatrix-calibrated)
`mmproj-gemma-4-E4B-it-Q8_0.gguf`	Multimodal projector (vision encoder + projection) — Q8_0
`vllm_config.yaml`	vLLM config (text-only path; kept for completeness, not viable for image input)
`config.json`, `processor_config.json`, `generation_config.json`, `tokenizer*`, `chat_template.jinja`	HF configs / tokenizer / chat template

Compression

The compression is fully llama.cpp-based and uses importance-matrix-guided 4-bit quantization to preserve quality at low bit-width:

F16 GGUF conversion. The original Gemma-4-E4B-it checkpoint is converted to a full-precision GGUF in F16 (convert_hf_to_gguf.py).
Importance-matrix (imatrix) computation. An imatrix is computed from a calibration dataset with llama.cpp's imatrix tool.
Imatrix-guided Q4_K_M quantization. The F16 GGUF is quantized to Q4_K_M with llama-quantize, passing the imatrix file so the K-quant mix uses the importance information.

The vision projector is shipped as Q8_0 — quantizing the small projector to 8-bit instead of carrying the BF16 file (~990 MB) saves bandwidth and VRAM without measurable quality loss.

Inference — `llama-server` (required for image input)

llama-server \
  -m gemma-4-q4_k_m.gguf \
  --mmproj mmproj-gemma-4-E4B-it-Q8_0.gguf

llama-server exposes an OpenAI-compatible /v1/chat/completions endpoint, uses the included chat_template.jinja automatically, and accepts images via the standard image_url content blocks.

The equivalent CLI for local testing is llama-mtmd-cli:

llama-mtmd-cli \
  -m gemma-4-q4_k_m.gguf \
  --mmproj mmproj-gemma-4-E4B-it-Q8_0.gguf \
  --image /path/to/image.jpg \
  -p "Describe the image."

vLLM (text-only — not used for evaluation)

A vllm_config.yaml is provided at the repo root and would be picked up by vllm serve …. However, vLLM's GGUF backend does not support multimodal Gemma 4: it can load the text model but cannot consume images. The image-to-text task therefore runs only under llama.cpp for this submission.

Hardware target

GPU: NVIDIA L4 (24 GB), single GPU.
Runtime: latest llama.cpp / llama-server (with multimodal --mmproj support, available since the Gemma 4 vision PR in llama.cpp).

License

Released under the Gemma license (https://ai.google.dev/gemma/terms), the same license as the base model.

Downloads last month: 39

GGUF

Model size

8B params

Architecture

gemma4

Hardware compatibility

4-bit

Model tree for Mathos34400/resilient-challenge-image-to-text

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Quantized

(196)

this model