Google's smallest Gemma 3 model at 4B parameters. Lightweight chat, copilots, coding and reasoning, cost-effective fine-tuned vertical assistants.
Gemma 3 4B is the compact member of Google's Gemma 3 open-model family, announced on March 12, 2025. It is a 4-billion-parameter vision-language model with image and text input, text output, a 128K-token context window, and the Gemma community's standard instruction-tuned variant for chat-style use. Google positions Gemma 3 as its most capable open family at launch, with longer context, multimodality, function calling, structured outputs, and coverage for more than 140 languages.
For TheRouter, Gemma 3 4B is a pragmatic open-weight choice when a team wants image understanding and multilingual chat at lower cost and latency than larger Gemma or frontier hosted models. It is more capable than tiny text-only models for document QA, visual inspection, extraction, and general assistant workloads, and is the practical Gemma 3 size when deployment footprint matters. It should not be treated as a Gemini replacement: the best fit is controlled low-to-mid-complexity work where predictable cost and open-weight portability matter more than peak reasoning.
Google publishes Gemma 3 4B as open weights for self-hosting, fine-tuning, and deployment across local hardware, Hugging Face, Kaggle, Google AI Studio, Vertex AI, and other runtimes.
TheRouter serves it as a managed OpenAI-compatible chat-completions route with text/image input and text output; this page does not claim local weight download, fine-tuning, or vendor-native runtime controls on the live TheRouter request path.
| Type | Rate |
|---|---|
| Input | $0.0432 / 1M tokens |
| Output | $0.0864 / 1M tokens |
| Release date | 2025-03-12Google Developers Blog ↗ | verified |
| Architecture / size | Dense transformer, 4B parametersGoogle Developers Blog ↗ | verified |
| Context window | Up to 128K tokensGoogle Developers Blog ↗ | verified |
| Training tokens | 4T tokens for the 4B modelGoogle Developers Blog ↗ | verified |
| Input / output modalities | Text and images in; text outGoogle Developers Blog ↗ | verified |
| License | Gemma Terms of Useai.google.dev ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU Pro Gemma 3 IT 4B, 0-shot, as reported in Google's Gemma 3 model card | 44.9% | ai.google.dev ↗ | |
HumanEval Gemma 3 IT 4B, 0-shot, code generation benchmark from the official model card | 54.3% | ai.google.dev ↗ | |
GSM8K Gemma 3 IT 4B, 0-shot, grade-school math benchmark from the official model card | 75.6% | ai.google.dev ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "google/gemma-3-4b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "google/gemma-3-4b",
"messages": [{"role": "user", "content": "Explain adaptive image tiling in two paragraphs."}]
}'Google's launch post presents Gemma 3 as a four-size open-model family, with 1B, 4B, 12B, and 27B variants, image-and-text input, text output, structured outputs, function calling, and more than 140 supported languages.
Use it when open weights, visual input, and predictable cost matter. For highest reasoning quality, longer context, or managed Google-native tools, choose a Gemini route instead.
No. Gemma 3 4B is a vision-language model that can read image input and return text. Use a dedicated image model for generation.
Google's release lets teams download and run weights in multiple environments. TheRouter exposes a managed OpenAI-compatible API route, so integration is simpler but local runtime and fine-tuning controls are outside this route.
No. This curator run has no paid API budget, so snippet verification remains unproven until a human operator approves and funds the verification run.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | Google Developers Blog ↗ | 2026-08-21 | verified |
| Architecture / size | Google Developers Blog ↗ | 2026-08-21 | verified |
| Context window | Google Developers Blog ↗ | 2026-08-21 | verified |
| Training tokens | Google Developers Blog ↗ | 2026-08-21 | verified |
| Input / output modalities | Google Developers Blog ↗ | 2026-08-21 | verified |
| License | ai.google.dev ↗ | 2026-08-21 | verified |
| MMLU Pro | ai.google.dev ↗ | 2026-08-21 | verified |
| HumanEval | ai.google.dev ↗ | 2026-08-21 | verified |
| GSM8K | ai.google.dev ↗ | 2026-08-21 | verified |
| Google introduces Gemma 3 with multimodality and 128K context | Google Developers Blog ↗ | 2026-08-21 | verified |