Text and chat · Google DeepMind

Gemma

Google's small open models, run in private mode: very cheap, they read images and video and call tools.

Private Low cost
Latest versionGemma 4April 2026
Artificial Analysis Intelligence Index17 points#71 of 90 on Artificial Analysis Intelligence Index
Price$0.162 per million tokens written$0.000113 a page written
Key spec256K tokens of context

What sets it apart

Gemma is Google DeepMind’s small open family built from the same research as Gemini. Gemma 4 moves to the permissive Apache 2.0 license, a 256K context, configurable thinking, and image and video input, in a 31B dense and a 26B mixture of experts with only 3.8B active.

  • Apache 2.0 from Gemma 4
  • 26B MoE with 3.8B active
  • 140+ languages in pre-training

Strengths and limits

Where it shines

  • Among the cheapest models that read images
  • Private mode, tool calling, 256K context (Gemma 4)

Think twice when

  • Small models: less reliable on long reasoning than large ones
  • Short replies: 8K tokens of output at most on Gemma 4
  • Gemma 4: 256K context, image and video input
  • Open weights

Use cases

A private assistant on your own hardware

Summarize this 20-page report in five bullet points and list any figure that looks inconsistent.

Gemma 4 has open Apache 2.0 weights, so it can run entirely offline.

Multilingual rewriting

Translate this product description into German, Spanish and Italian, keeping the tone friendly.

Google reports 140+ languages in pre-training and 35+ with native support.

Against its rivals

  • Qwen open modelsQwen3.8 27B scores 34 on the index against 15 for Gemma 4 31B.#23 of 90 on Artificial Analysis Intelligence Index (40 points)
  • GPT OSSgpt-oss-120b is larger, text only, and scores 12 against 15.#79 of 90 on Artificial Analysis Intelligence Index (12 points)
  • DeepSeekDeepSeek is far larger and stronger but needs data-center hardware.#24 of 90 on Artificial Analysis Intelligence Index (39 points)
Compare side by side ↗

Version history

Each version, dated, with what it changed over the one before.

Artificial Analysis Intelligence Index, version after versionHigher is better · Artificial Analysis

Hollow points are earlier versions no longer offered in Sub Rosa.

See the numbers
VersionReleasedpoints
Gemma 3 27B2025-035
Gemma 42026-0417
  1. Gemma 3 27B
    • 27.4B dense model with a 128K window and image input, under the Gemma license.
    Source : artificialanalysis.ai
  2. Gemma 4Latest
    • Apache 2.0 license, 256K context and configurable thinking modes, with native function calling and system prompts.
    • Adds video input; the 26B variant is a mixture of experts with 3.8B active parameters.
    Source : huggingface.co

Benchmarks

How to read these scores: what a benchmark tells you.

Artificial Analysis Intelligence Index: #71 of 90 models measured in the catalog. One score averaging about ten independent tests of reasoning, knowledge, math, coding and agent work, all run the same way for every model. Higher is better. Each model is listed at its highest reasoning effort. Two or three points apart is a close call; ten points is a different class of model.

Artificial Analysis Intelligence IndexHigher is better · Artificial Analysis, October 4, 2026
  1. 1. Claude Opus 5.5Claude58
  2. 2. Claude Sonnet 5.5Claude56
  3. 3. Claude Fable 5.1Claude53
  4. 4. GPT-6 AstraGPT53
  5. 5. Gemini 4 ArgonGemini53
  6. 6. GPT-6.1 SolGPT52
  7. 7. Muse WonderMuse52
  8. 8. Claude Opus 5Claude51
  9. 9. Claude Fable 5Claude50
  10. 10. GPT-6 SolGPT48
  11. 71. Google Gemma 4 26B A4B InstructGemma17
  12. 75. Google Gemma 4 31B InstructGemma15
  13. 90. Google Gemma 3 27B InstructGemma5
See the numbers
Modelpoints
1. Claude Opus 5.558
2. Claude Sonnet 5.556
3. Claude Fable 5.153
4. GPT-6 Astra53
5. Gemini 4 Argon53
6. GPT-6.1 Sol52
7. Muse Wonder52
8. Claude Opus 551
9. Claude Fable 550
10. GPT-6 Sol48
71. Google Gemma 4 26B A4B Instruct17
75. Google Gemma 4 31B Instruct15
90. Google Gemma 3 27B Instruct5

Announced by the maker

Measured by the maker with its own settings: useful, but not directly comparable with the independent scores above.

BenchmarkModelScoreSource
gpqa-diamondGoogle Gemma 4 31B Instruct84.3huggingface.co

All 3 models in this family

ModelUsePrivacyPrice
Google Gemma 4 31B Instructgoogle-gemma-4-31b-itTextPrivate$0.162 per million tokens written$0.000113 a page written
Google Gemma 4 26B A4B Instructgoogle-gemma-4-26b-a4b-itTextPrivate$0.18 per million tokens written$0.000126 a page written
Google Gemma 3 27B Instructgoogle-gemma-3-27b-itTextPrivate$0.0899 per million tokens written$0.0000629 a page written

Specifications

SpecificationGemma 4 31BGemma 4 26B A4BGemma 3 27B
ReleasedApril 2026April 2026March 2025
Parameters30.7B25.2B total, 3.8B active27.4B
WeightsOpen, Apache 2.0Open, Apache 2.0Open, Gemma license
Acceptstext, images, videotext, images, videotext, images
Producestexttexttext
Context256K tokens256K tokens198K tokens
Languages140+ languages in pre-training··
AbilitiesReads images, Uses Sub Rosa’s tools, Thinks before answeringReads images, Uses Sub Rosa’s tools, Thinks before answeringReads images, Uses Sub Rosa’s tools
Price$0.162 per million tokens written$0.18 per million tokens written$0.0899 per million tokens written
PrivacyPrivatePrivatePrivate

Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.