{"slug":"gemma-3","id":"gemma-3","type":"model","title":"Gemma 3","description":"Google's open model family optimized for on-device and edge deployment. Multimodal from 4B parameters. Sizes from 1B to 27B.","last_updated":"2026-04-10","last_verified":null,"verification_status":"unverified","markdown_url":"/content/models/gemma-3.md","html_url":"/models/gemma-3","api_url":"/api/v1/models/gemma-3.json","content_hash":"9b8c41f767f8ab3f20fb6a19336914199798ee25d1904f377f75061cfd00fb3f","sha256":"9b8c41f767f8ab3f20fb6a19336914199798ee25d1904f377f75061cfd00fb3f","provider":"Google","pricing":{"input":"Free (open weights)","output":"Free (open weights)","free":true},"benchmarks":{"reasoning":75,"coding":73,"math":72,"writing":76,"multilingual":78,"speed":93},"tags":["google","open-source","text","image"],"website":"https://ai.google.dev/gemma","release_date":"2025","relationships":{"links":[],"related":[{"id":"gemma-4","title":"Gemma 4","type":"model","html_url":"/models/gemma-4","markdown_url":"/content/models/gemma-4.md","shared_tags":["google","open-source","text","image"],"score":8},{"id":"gemini-3-flash","title":"Gemini 3 Flash","type":"model","html_url":"/models/gemini-3-flash","markdown_url":"/content/models/gemini-3-flash.md","shared_tags":["google","text","image"],"score":7},{"id":"gemini-3.1-pro","title":"Gemini 3.1 Pro","type":"model","html_url":"/models/gemini-3.1-pro","markdown_url":"/content/models/gemini-3.1-pro.md","shared_tags":["google","text","image"],"score":7},{"id":"llama-4-maverick","title":"Llama 4 Maverick","type":"model","html_url":"/models/llama-4-maverick","markdown_url":"/content/models/llama-4-maverick.md","shared_tags":["open-source","text","image"],"score":5},{"id":"llama-4-scout","title":"Llama 4 Scout","type":"model","html_url":"/models/llama-4-scout","markdown_url":"/content/models/llama-4-scout.md","shared_tags":["open-source","text","image"],"score":5},{"id":"mistral-small-4","title":"Mistral Small 4","type":"model","html_url":"/models/mistral-small-4","markdown_url":"/content/models/mistral-small-4.md","shared_tags":["open-source","text","image"],"score":5}],"explicit":{}},"metadata":{"title":"Gemma 3","type":"model","id":"gemma-3","provider":"Google","model_type":"open-source","release_date":"2025","description":"Google's open model family optimized for on-device and edge deployment. Multimodal from 4B parameters. Sizes from 1B to 27B.","last_updated":"2026-04-10","context_window":"128K tokens","website":"https://ai.google.dev/gemma","license":"Gemma Terms of Use","modality":["text","image"],"tags":["google","open-source","text","image"],"pricing":{"input":"Free (open weights)","output":"Free (open weights)","free":true},"benchmarks":{"reasoning":75,"coding":73,"math":72,"writing":76,"multilingual":78,"speed":93},"parameters":"1B to 27B variants","hardware_requirements":"27B: 16GB VRAM; 4B: runs on phones; 1B: runs on embedded devices","best_for":["On-device AI","Mobile apps","Edge deployment","IoT","Low-resource environments"]},"content_text":"# Gemma 3\n\nThe model that made on-device AI real. Gemma 3 brought multimodal capability down to 4B parameters -- small enough to run on a phone -- while the 1B variant fits on embedded devices. Google optimized these models for the edge first and benchmarks second, and it shows: speed at 93/100 is near the top of anything in our rankings.\n\nThe benchmark scores are honest about what you get at this size. Reasoning at 75, coding at 73, math at 72 -- none of these compete with the big MoE models. But that is not the point. The point is that you get a multimodal model with 128K context that runs on 16GB of VRAM at the largest (27B) variant, or on a phone at 4B. No API calls, no latency, no data leaving the device.\n\nCompared to Phi-4 at a similar size class, Gemma 3 trades slightly lower coding scores for better multilingual support (78 vs 72) and native multimodal capability that Phi-4 lacks entirely. The Gemma Terms of Use license is more restrictive than Apache 2.0 but still allows commercial use.\n\nWith Gemma 4 now released, Gemma 3 is the previous generation. It remains relevant for deployments where the newer model's hardware requirements are too high, particularly the 1B and 4B tiers that Gemma 4 does not fully replace.\n\n**When to pick something else:** Gemma 4 is strictly better if your hardware supports it. For a small model that prioritizes coding, Phi-4 edges ahead. For multilingual edge deployment, Cohere Tiny Aya covers 70+ languages at a similar size.","content_length":2355,"generated_at":"2026-04-24"}