Gemma 3
Google's Open-Weight Family
Google DeepMind's March 2025 open-weight family, sized 1B / 4B / 12B / 27B. Gemma 3 added vision input (from 4B up), 128K context, and 140-language coverage — making it the most capable single-GPU open model from a major US lab at release.
Intelligence
Below avg
Speed
Slow
29 tok/s output
Cost
Low
$0.04 in / $0.08 out
Context
128K
Up to 128,000 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Made Google a real player in the open-weight tier — until Gemma 3, Llama and Qwen had owned that space outright. The vision support and 128K context closed key feature gaps with Llama 3.2 Vision and Qwen 2.5-VL.
Core Capabilities
Long Documents
Handles entire codebases, books, and multi-doc RAG.
Multimodal
Combines text, vision, and audio in one model.
Generative
Produces images, video, audio, or other media.
Agent Workflows
Built for tool use and autonomous tasks.
Context Window
128k tokens
≈ 98 pages
4k Chat 聊天
32k Long docs 长文档
128k This model 本模型
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
Availability
API
Not available
Product / App
Not available
Open Source
Released
Enterprise
—
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Context / memory
Context window size · log-scaled
6.0
9.0
6.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- Four sizes (1B / 4B / 12B / 27B) — covers laptop-class up to single-GPU server inference
- Gemma-3-4B-IT beats Gemma-2-27B-IT; Gemma-3-27B-IT beats Gemini 1.5 Pro on shared benchmarks
- Vision input from 4B up; 1B is text-only — first Gemma family with multimodal at small sizes
- 128K context window across all sizes; 140+ language coverage
- Interleaves 5 local-attention layers per global layer — saves memory at long context
- Distillation + RL + model merging in post-training delivered noticeable math/coding/instruction gains
Best use cases
- Single-GPU production inference (27B fits H100 / A100 80GB at FP precision)
- Laptop / phone deployment via the 1B / 4B variants
- Multilingual apps in 140 languages at open-weights pricing
- Vision + text workflows that don't justify a frontier API call
Tools to try
Not ideal for
- Frontier reasoning leaderboards — Gemini 2.5 / 3 Pro and competitors lead
- Workloads needing 1M+ context (Gemma 3 caps at 128K)
- Audio or video input (Gemma 3 base is text + image)