gemma-3-4b-it
Q4 GGUF download: google/gemma-3-4b-it-qat-q4_0-gguf - 3.16 GB on disk, read 2026-11-16. VRAM figures above are estimates; this is the actual file size.
Hosted API cost & speed
What it costs to rent this model instead of running it locally
| Metric | Value |
|---|---|
| Input (per 1M tokens) | $0.00 |
| Output (per 1M tokens) | $0.00 |
Median across hosted providers, measured by Artificial Analysis - updated 2026-11-21. Local llama.cpp speed depends on your hardware — see the hardware pages for estimates.
Arena rating (LMArena)
Human preference wins from head-to-head chat battles
| Metric | Value |
|---|---|
| Overall arena rating | 1290.8 |
| Rank (overall) | #266 |
| Battle votes | 4,171 |
Overall rating published by LMArena on 2026-11-19 - a pairwise human-preference Elo across live chat battles; higher is better. Fetched 2026-11-21. Open-weights models compete head-to-head with proprietary ones here — no benchmark prompts, just what people pick.
Will it run on your hardware?
Pick your GPU memory - see which quantizations fit, and the cheapest card for the rest
Need an exact figure for your context length? Use the VRAM calculator.
Run it locally
Copy-paste - running in under a minute
vllm serve google/gemma-3-4b-itNew to this? Start with Ollama · serve to many users with vLLM.
Deep dive
Notes, sources, and the full write-up
gemma-3-4b-it is an open-weight vision-language model published by google on Hugging Face in February 2025, with 4.3B parameters under the gemma license. It is built on google/gemma-3-4b-pt.
Among the 74 vision-language models in this directory it ranks 20th by monthly Hugging Face downloads, and 62 of the others are larger.
HardwareLink to this section
Unquantized FP16 weights take about 9 GB; 8-bit brings that to ~5 GB and 4-bit (Q4_K_M) to ~2 GB, before activations and any context or image buffers. A 12 GB consumer card holds it at 4-bit; FP16 needs a 24 GB card such as an RTX 4090. Check exact numbers for your card with the VRAM calculator.
Similar vision-language modelsLink to this section
- Qwen3.5 4B - 4.7B, 4,416,216 downloads/month
- Nanonets-OCR2-3B - 3.8B, 898,672 downloads/month
- Kromoss - 5.1B, 62,589 downloads/month
- VideoChat3-4B - 4.5B, 2,069 downloads/month
- TimeLens2-4B - 4.4B, 1,076 downloads/month
PopularityLink to this section
gemma-3-4b-it was downloaded 1,007,407 times on Hugging Face in the last month (28,435,396 all-time) and has 1,694 likes.
Run it locallyLink to this section
Weights and usage instructions are on the google/gemma-3-4b-it model card; it loads with Hugging Face Transformers.
Local runtimes that fit this model:
- vllm - high-throughput serving on NVIDIA GPUs
- transformers - the reference Python library most model cards target
PaperLink to this section
Frequently asked
Quick answers to common questions
How much VRAM does gemma-3-4b-it need?
gemma-3-4b-it with 4.3B parameters needs approximately 2 GB at Q4_K_M quantization. Use our VRAM calculator for an exact estimate.
Is gemma-3-4b-it better than other google models?
gemma-3-4b-it has 4.3B parameters with 8,192 context - a strong choice for vision.
What license is gemma-3-4b-it under?
gemma-3-4b-it is released under the gemma license, making it suitable for most commercial and personal projects.
What hardware runs gemma-3-4b-it well?
With 4.3B parameters, gemma-3-4b-it requires adequate VRAM. High-end GPUs like the RTX 4090 (24GB), RTX 5090 (32GB), or Mac Studio with unified memory are good options. Check our hardware directory for specific recommendations.
What is the best quantization for gemma-3-4b-it?
Q4_K_M is the recommended sweet spot - ~98% of FP16 quality at ~27% of the size. Q5_K_M (~3 GB) is an option if you have spare VRAM. Use our VRAM calculator to compare.
What models compete with gemma-3-4b-it?
gemma-3-4b-it competes with other models in its class. Browse our model directory for comparisons, benchmarks, and community reviews to find the best fit.
Nearby options
Similar models and compatible hardware by spec