- Size
- 4.3B parameterscounted from the safetensors header on the Hub
- Licence
- unknown, Gemma Terms of UseOur reading, check the licence: Commercial use is allowed with no user or revenue cap, but the Gemma Prohibited Use Policy (a separate document, not in this text) is part of the terms, and you must pass the use restrictions on as enforceable terms to anyone you distribute to. Models distilled from Gemma or trained on its outputs to perform like it count as Model Derivatives and stay under these terms, and Google may restrict use remotely if it believes the terms are broken. The page says these terms cover the models in its appendix (Gemma 1 to 3, 3n and variants) and that Gemma 4 has a separate licence.
- On the Hub since
- 20 Feb 2025the date the repo was created
- Context length
- 128k tokensas an API provider reports it
- Fine-tune
- of google/gemma-3-4b-ptGemma3ForConditionalGeneration
- On Hugging Face
- google/gemma-3-4b-it985k downloads in the last 30 days, 1.7k likes
Scores
- default setting23.2%±1.928 Aug 2026, log
Epoch's name: gemma-3-4b-it. Best closed model on this set: gpt-6.2-astra_max, 96.8% (OpenAI). All 341 results
- default setting7.5%±2.828 Aug 2026, log
Epoch's name: gemma-3-4b-it. Best closed model on this set: claude-fable-5_high, 100.0% (Anthropic). All 318 results
- default setting50.9%conceptualreasoning.ai
Epoch's name: gemma-3-4b-it. Best closed model on this set: claude-fable-5-2_unknown, 99.2% (Anthropic). All 241 results
- default setting2.8%conceptualreasoning.ai (error bar left out)
Epoch's name: gemma-3-4b-it. Best closed model on this set: claude-fable-5-2_xhigh, 70.6% (Anthropic). All 203 results
Claimed by the maker Claimed
The model card has no claimed scores in a machine-readable form (the card's model-index), so none are shown. The maker may give scores in the card text: read the card.
Other views, which we link to and do not copy: look this model up on Artificial Analysis and the Arena leaderboard.
Independent scores from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Each row links its own source. What the badges mean.
Run it yourself
Memory needed = weights + context cache + 0.50 GB for the runtime. Context cache at 8k tokens: 0.27 GB, worked out from the model's config (dongchengli/gemma-3-4b-it-text-int4-gptq), at 16-bit precision. Choose a machine above to see a verdict for each format. Method
GGUF for llama.cpp, Ollama, LM Studio and others
google/gemma-3-4b-it-qat-q4_0-gguf (the maker's own conversion), 5.6k downloads. Sizes are the real files on the Hub, all parts added up. For image input you also need the projector file mmproj-model-f16-4B.gguf (0.79 GB), not counted below.
| Format | Size | Memory needed | Get it |
|---|---|---|---|
| Q4_0 | 2.9 GBfile size | 3.7 GB2.9 + 0.27 + 0.50 | Commandsllama.cpp llama-server -hf google/gemma-3-4b-it-qat-q4_0-gguf:Q4_0Ollama ollama run hf.co/google/gemma-3-4b-it-qat-q4_0-gguf:Q4_0Download hf download google/gemma-3-4b-it-qat-q4_0-gguf --include "gemma-3-4b-it-q4_0.gguf" --local-dir gemma-3-4b-it-qat-q4_0-gguf |
Original weights size from the safetensors header, for vLLM, SGLang, Transformers
| Format | Size | Memory needed | Repo |
|---|---|---|---|
| BF16original weights | 8.0 GBfile size | 8.8 GB8.0 + 0.27 + 0.50 |
hf download google/gemma-3-4b-it --local-dir gemma-3-4b-itOther conversions (12; sizes not read)
- unsloth/gemma-3-4b-it-unsloth-bnb-4bit bitsandbytes, 4bit, 19k downloads
- unsloth/gemma-3-4b-it-bnb-4bit bitsandbytes, 4bit, 4.9k downloads
- dongchengli/gemma-3-4b-it-text-int4-gptq GPTQ, int4, 3.9k downloads
- onnx-community/gemma-3-4b-it-ONNX Other, 1.2k downloads
- RedHatAI/gemma-3-4b-it-quantized.w4a16 Other, w4a16, 1.1k downloads
- positron-ai/google_gemma-3-4b-it-ingest-best-gptq GPTQ, 990 downloads
- gaunernst/gemma-3-4b-it-int4-awq AWQ, int4, 840 downloads
- RedHatAI/gemma-3-4b-it-FP8-dynamic FP8, 791 downloads
- ISTA-DASLab/gemma-3-4b-it-GPTQ-4b-128g GPTQ, 703 downloads
- RedHatAI/gemma-3-4b-it-quantized.w8a8 Other, w8a8, 658 downloads
- pytorch/gemma-3-4b-it-HQQ-INT8-INT4 Other, int8, 364 downloads
- unsloth/gemma-3-4b-it-FP8-Dynamic FP8, 189 downloads
The Hub tags 100 or more quantised repos as conversions of this model; we read the most downloaded and show the maker's own first.
Run it by API
3 providers serve this model through the Hugging Face router. Prices are US dollars per million tokens. Precision is what the provider reports; a lower precision than the original weights can change answers. Live first, then cheapest output.
| Provider | In | Out | Context | Precision | Tokens/s | First token | Uptime 24h | Tools | JSON |
|---|---|---|---|---|---|---|---|---|---|
| Parasail | $0.04 | $0.05 | 128k | bf16 | 79 | 0.9 s | 99.9% | no | yes |
| DeepInfra | $0.05 | $0.10 | 128k | bf16 | 56 | 1.0 s | 100.0% | no | yes |
| featherless-ai | ? | ? | ? | not given | ? | ? | ? | not given | not given |
Through the router, with a Hugging Face token (the router bills your Hugging Face account). Change the name after the colon to pick another provider.
curl https://router.huggingface.co/v1/chat/completions \
-H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" \
-d '{"model": "google/gemma-3-4b-it:parasail", "messages": [{"role": "user", "content": "Hello"}]}'Not built yet: one key for every open model, with BlockRun, and our own tested-quality score per provider. Until then, precision and measured speed are the best guide to a provider here.
Fine-tunes of this model
Repos that declare this model as their base, most downloaded first. Listing is not a recommendation; unreviewed ones have not been checked.
- unsloth/gemma-3-4b-it4.3B50kUnreviewed
- nvidia/Nemotron-3-Content-Safety4.3B11kUnreviewed
- nvidia/Nemotron-Content-Safety-Reasoning-4B4.3B2.8kUnreviewed
- google/gemma-3-4b-it-qat-q4_0-unquantized4.3B85Unreviewed
- google/gemma-3-4b-it-qat-int4-unquantized4.3B35Unreviewed
Planned for this page Not built yet
- Try it in private chat
- A link to Solana AI's confidential chat for the models it serves. This model is not listed as served there.
- Fine-tune it
- Fine-tuning a few base models from this page and publishing the result here.
- Creators and earnings
- Pages for people who publish fine-tunes and adapters of this model, and what they earn.
- Fund and bounties
- Bounties for new open models, paid from a public fund. How the fund will work.
- Contribute compute
- How to give compute to open model work, with Akash.
Model data from the Hugging Face Hub 19 Nov 2026, 13:43 UTC. The repo. Found a mistake on this page? Tell us.