12
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

ローカルLLM:「Gemma 4」を LM Studio で少し動かしてみる(環境は メモリ32GB の M5 の MacBook Pro)

12
Last updated at Posted at 2026-04-06

はじめに

2026/4/3 に発表された「Gemma 4」に関する記事です。

以下のポストにあるように、LM Studio で使えるようになっているので、軽く動作確認をやってみました。

関連情報など

モデルの情報

冒頭でポストされていた内容を引用すると、Gemma 4 は、E2B、E4B、26B A4B、31B という種類があるようです。

それについて、LM Studio公式の「Model Catalog」を見てみると、冒頭に Gemma 4 のリンクが出てきています(※記事執筆時点)。

image.png

以下は、Gemma 4 を選択した後の画面です。

https://lmstudio.ai/models/gemma-4
2026-04-06_12-26-07.jpg

対応している機能・ライセンス

モデル名の右にあるマークから、「画像入力」「Tool use」「Reasoning」に対応したモデルとなっているようです。

2026-04-06_12-35-22.jpg

それと、X で言及されているポストを複数見かけた気もする話ですが、Gemma 4 のライセンスは(Gemma 3 とは異なり)「Apache 2.0」になっているようです。

自分が試すモデル

自分が試すモデルは、この記事を書く準備をした時点で LM Studio上でモデルリスト上のトップに出てきたGoogle が提供するモデルの GGUF版の 1つ(google/gemma-4-26b-a4b)です。
※ LM Studio のコミュニティも、lmstudio-community/gemma-4-26B-A4B-it-GGUF を提供しているようですが

2026-04-04_17-42-45.jpg

自分が試す環境

今回、自分が Gemma 4 を試す環境は、M5 の MacBook Pro(メモリ 32GB)です。

VRAM として、以下(※ LM Studio の設定画面)のように、24.96 GB を利用できる環境です。

2026-04-06_12-39-46.jpg

今回用いるモデルは、VRAM のみで十分利用できるサイズとなります。

軽く動作確認をやってみる

簡単な処理を 2種類試す

とりあえず軽く動作確認をやってみます。

試す内容は、適当なもの(「あなたは誰?」という簡単なプロンプト)です。Gemma 4 は Reasoning 対応のモデルで、LM Studio上で Reasoning の ON/OFF ができるため、その 2パターンは試すことにします。
※ プロンプトを入力する部分の「Think」と書かれたところをクリックして、Reasoning の ON/OFF を切り替えられます

試した結果

試した結果は、以下のとおりです。シンプルなプロンプトなので、回答もシンプルです。

2026-04-04_18-34-55.jpg

処理にかかった時間などは、以下となりました。

  • Reasoning ON
    • Reasoning の時間: 8.21秒(「Thought for 8.21 seconds」という表示の部分が該当)
    • トークンの出力スピード: 約 38 トークン/秒
  • Reasoning OFF
    • トークンの出力スピード: 約 42 トークン/秒

体感としては、わりと良い感じのスピードでした。

試した結果の補足: Reasoning ありの場合の内容

それと、以下は Reasoning の過程として出力されたものです。

2026-04-04_18-42-32.jpg

テキスト化した内容を、折りたたみで掲載しておきます。

Reasoning の内容

User asks: "あなたは誰?" (Who are you?)
Language: Japanese.

Name: Gemma 4.

Developer: Google DeepMind.

Nature: Large Language Model (LLM).

Type: Open weights model.

Capabilities: Process text and images (audio for specific versions), output text only. No image generation. Knowledge cutoff Jan 2025.

State clearly that I am Gemma 4.

Mention Google DeepMind as the developer.

Describe myself as an LLM with open weights.

Maintain a helpful and polite tone in Japanese.

Draft 1 (Internal Monologue): 私はGemma 4です。Google DeepMindによって開発された大規模言語モデルです。テキストと画像を理解できますが、出力はテキストのみです。

Refinement (Polishing for natural Japanese): 「私はGemma 4です。Google DeepMindによって開発された、オープンウェイトの大規模言語モデル(LLM)です。」

"I am Gemma 4." -> 「私はGemma 4です。」

"Developed by Google DeepMind." -> 「Google DeepMindによって開発されました。」

"Large Language Model with open weights." -> 「オープンウェイトの大規模言語モデル(LLM)です。」

(Ensuring no claims of image generation or unauthorized tool use).

私はGemma 4です。Google DeepMindによって開発された、オープンウェイトの大規模言語モデル(LLM)です。

12
11
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
12
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?