Gemma 3 4B

Tool callingStructured outputImage input

Model card Hugging Face Launch post Website

Open multimodal Gemma instruction model for efficient text generation and image understanding

Context: 131K
Released: 1y ago
Parameters: 4B
Input / 1M tokens: $0.05
Output / 1M tokens: $0.10

Sources:models.devOpenRouter

Benchmarks

DeepSWE is the primary coding benchmark where an exact model evaluation exists. Other sourced evaluations appear below; missing scores are not zero.

Not evaluated in DeepSWE 1.1 as of 2026-09-22. Results are only shown for exact model versions in the official dataset. Official leaderboard.

Other published benchmarks

Evaluation	Reported score and source
GPQA diamond	23.2% Epoch AI · gemma-3-4b-it Epoch AI evaluation
OTIS Mock AIME 2024-2025	7.5% Epoch AI · gemma-3-4b-it Epoch AI evaluation

Benchmark compilation: Epoch AI, Capabilities & benchmarking (CC BY attribution). Scores are converted to percentages and model names matched to the Vertical Tabs catalog. Individual evaluations are credited to their original sources. External results may carry separate licences and evaluation configurations; do not interpret scores from different harnesses as directly comparable.

Specifications

Developer: Google DeepMind
Model family: Gemma
Type: Language model
Predecessor: —
Successor: —
Released: March 12, 2025
Model ID: google/gemma-3-4b-it
Openness: Open weights
License: —
Parameters: 4B
Input modalities: text, image
Output modalities: text
Modality: text+image->text
Native context: 128K
Tokenizer: Gemini
Max output tokens: 16K
Knowledge cutoff: 2024-08
Hugging Face: google/gemma-3-4b-it
Source freshness: Metadata synced Oct 8, 2026, 10:17 PM UTC

Additional source evidence

Source model: google/gemma-3-4b-it
Source family: gemma
Last updated: 2025-03-12
Knowledge cutoff: 2024-08
Model context: 131K
Model output limit: 131K
Open weights: Yes
File attachments: Yes
Reasoning: No
Tool calling: Yes
Temperature control: Yes

Downloadable weights

Hugging Face

Complete source record (JSON)

{
  "id": "google/gemma-3-4b-it",
  "name": "Gemma 3 4B IT",
  "description": "Open multimodal Gemma instruction model for efficient text generation and image understanding",
  "family": "gemma",
  "last_updated": "2025-03-12",
  "knowledge": "2024-08",
  "attachment": true,
  "reasoning": false,
  "tool_call": true,
  "temperature": true,
  "modalities": {
    "input": [
      "text",
      "image"
    ],
    "output": [
      "text"
    ]
  },
  "limit": {
    "context": 131072,
    "output": 131072
  },
  "release_date": "2025-03-12",
  "open_weights": true,
  "weights": [
    {
      "label": "Hugging Face",
      "url": "https://huggingface.co/google/gemma-3-4b-it"
    }
  ]
}

Infrastructure

Provider availability

All providers

Provider	Endpoint	Context	Input /M tokens	Output /M tokens	Cached input /M tokens	Cache write /M tokens	Uptime 1d	Data handling
Parasail	parasail/fp8	131K	$0.04	$0.05	$0.02	—	99.88%	ZDR: Yes Training: Unknown Prompt logging: Unknown Retention: Unknown ZDR source Privacy policy Checked 2026-10-15
DeepInfra	deepinfra/bf16	131K	$0.05	$0.10	—	—	100.00%	ZDR: Yes Training: Unknown Prompt logging: Unknown Retention: Unknown ZDR source Privacy policy Checked 2026-10-15

Provider data:OpenRouter

Text provenance

Gemini app/web

Google SynthID text watermarking

Google DeepMind says text generated in the Gemini app and web experience is watermarked with SynthID by adjusting token probabilities.

Official explanation

Deployment graph

Can I run this?

Vertical Tabs only shows compatibility when it has structured evidence tying this exact model to hardware, runtime, quantization, memory use, and observed performance.

Browse hardware