DeepMindオープンソース

Gemma 4 31B（稠密旗舰模型）

Name: Gemma 4 31B（稠密旗舰模型）
Author: DeepMind

このモデルを比較

Google DeepMind開発の高性能オープンソースモデル。31Bパラメータのdense architecture。

シェア:X はてブ LINE

パラメータ

コンテキスト長

256K

ライセンス

Apache 2.0

リリース日

2026-04-02

日本語性能

🌐多言語対応

一般的な多言語対応モデル。基本的な日本語処理は可能だが、特化モデルには劣る。

API料金

このモデルのAPI料金情報は現在未公開です

強み

弱み

活用例

深度分析

Arena Elo

1451

#3 open dense model overall

MMLU Pro

85.2%

vs Qwen3.5-27B: 86.1%

AIME 2026 (no tools)

89.2%

vs Qwen3.5-27B: 84.2%

LiveCodeBench v6

80.0%

vs Qwen3.5-27B: 80.7%

GPQA Diamond

84.3%

vs Qwen3.5-27B: 85.5%

Codeforces ELO

2150

vs Qwen3.5-27B: 1899

強み

・Top-tier reasoning and math performance (89.2% on AIME 2026, ELO 2150 on Codeforces)
・Native function calling and thinking mode enable advanced agentic workflows
・Multimodal with 256K context and 140+ language support under permissive Apache 2.0 license

弱み

・Inference speed reported as slower than competitors (~11 tokens/sec on 26B MoE vs 60+ for Qwen)
・No native audio support on the 31B dense variant (limited to E2B, E4B, 12B models)
・Training data cutoff is January 2025, requiring RAG for recent knowledge

競合比較

Model	Arena	SWE	GPQA	Price
Gemma 4 31B	1451	41.6%	84.3%	Free (Open Weight)
Qwen3.5-27B	1444	72.4%	85.5%	Free (Open Weight)
GLM-5	1457	-	-	API Only

概要

Gemma 4 31B is Google DeepMind's flagship open-weight, dense Transformer model released under the Apache 2.0 license. It represents a significant leap over its predecessor Gemma 3 27B, achieving state-of-the-art performance for its size class in reasoning, coding, and agentic tasks. The model's standout features include native function calling, a configurable 'thinking' mode for step-by-step reasoning, and a 256K token context window. Built on the same research stack as Gemini 3, it achieves a strong balance between capability and accessibility, running on a single 80GB H100 or quantized on consumer hardware. Positioned as a versatile base for fine-tuning and deployment, Gemma 4 31B competes directly with models like Qwen3.5-27B. While it leads in human preference rankings (Arena Elo) and competitive programming (Codeforces), it slightly trails in some knowledge and agentic benchmarks. Its Apache 2.0 license eliminates previous usage restrictions, making it a compelling choice for commercial applications. However, community feedback highlights its slower inference speeds and lack of audio support in the dense variant as notable drawbacks compared to some competitors.

ベンチマーク＆性能

Gemma 4 31B demonstrates exceptional performance, particularly in reasoning and coding. Key benchmark scores from the official model card and technical report are: | Benchmark | Gemma 4 31B | Gemma 3 27B (No Think) | Qwen3.5-27B* | Notes | |-----------|-------------|------------------------|--------------|-------| | MMLU Pro | 85.2% | 67.6% | 86.1% | General knowledge | | AIME 2026 (no tools) | 89.2% | 20.8% | 84.2% | Advanced math reasoning | | LiveCodeBench v6 | 80.0% | 29.1% | 80.7% | Competitive programming | | Codeforces ELO | 2150 | 110 | 1899 | Algorithmic problem-solving | | GPQA Diamond | 84.3% | 42.4% | 85.5% | Graduate-level science QA | | τ²-Bench (Retail) | 86.4% | 6.6% | - | Agentic tool use | | HLE (with search) | 26.5% | - | - | Hard language understanding | | MMMU Pro | 76.9% | 49.7% | - | Multimodal understanding | | MATH-Vision | 85.6% | 46.0% | - | Visual math reasoning | | MRCR v2 (128k) | 66.4% | 13.5% | - | Long-context needle-in-haystack | *Qwen3.5-27B scores from DataLearner comparison page. The model shows massive generational gains, especially in math (AIME: 20.8% → 89.2%) and coding (LiveCodeBench: 29.1% → 80.0%).

詳細比較

### vs Qwen3.5-27B - **Context Window**: Gemma 4 31B offers 256K vs Qwen3.5's 1M tokens. - **Multimodality**: Gemma 4 31B supports video input (up to 60s) natively; Qwen3.5 does not. - **Coding & Reasoning**: Gemma leads on Codeforces (2150 vs 1899) and AIME (89.2% vs 84.2%). Qwen leads on SWE-bench Verified (72.4% vs 41.6%) and agentic tasks (τ²-Bench: 79.0 vs 76.9). - **Licensing**: Both use Apache 2.0. - **Speed**: Community reports suggest Qwen 3.5 has significantly faster inference (~60 tokens/sec vs ~11 tokens/sec on 26B MoE). ### vs GLM-5 (744B MoE) - **Scale**: GLM-5 is a much larger MoE model (744B total, 40B active). - **Performance**: GLM-5 has a slightly higher Arena Elo (1457 vs 1451). Its performance on reasoning and coding benchmarks is not directly comparable from the provided data. - **Accessibility**: GLM-5 is available as an API only, while Gemma 4 31B is open weight and self-hostable. - **Pricing**: Gemma is free to self-host; GLM-5 API pricing is undisclosed in the results. ### General Positioning Gemma 4 31B is the leading *dense* open model in its parameter class. It is optimized for reasoning and agentic workflows with native tool use. Compared to much larger MoE models like GLM-5 or Kimi K2.6, it offers superior performance-per-parameter and deployment flexibility on accessible hardware.

コミュニティ評価

Initial community reaction is mixed but positive overall. Developers and researchers have praised: 1. **The Apache 2.0 License**: Widely seen as the most important change, removing legal friction for commercial use. 2. **Multilingual Performance**: Early tests show strong results in non-English languages like German, Arabic, and Vietnamese. 3. **Coding & Reasoning Capabilities**: The leap in benchmarks is being validated in practical coding agent setups. However, notable concerns have emerged: 1. **Inference Speed**: The 26B MoE variant is reported as significantly slower (~11 tokens/sec) than competitors like Qwen 3.5, which is a major issue for production use. 2. **Missing Larger Model**: The rumored 124B MoE model was absent at launch, leaving a gap at the high end. 3. **Tooling Maturity**: While native function calling is praised, integration with existing agent frameworks (e.g., llama.cpp, Ollama) was reported as needing polish at launch. The Hugging Face team gave it a positive pre-release review, and it has gained traction for fine-tuning use cases.

ユースケース

**1. Local Coding Agent & IDE Assistant:** The model's high Codeforces ELO (2150) and native function calling make it ideal for powering offline coding assistants in tools like Cursor, VS Code, or Codex. It can handle complex algorithmic problems and interact with local file systems and APIs. Choose it over Qwen3.5 for competition-style algorithm problems. **2. Agentic Workflow Orchestration:** With configurable thinking mode and τ²-Bench score of 86.4%, it excels in multi-step tool-use tasks. For example, it can plan a research workflow that searches the web, calculates figures, and translates results. Prefer it over larger models when latency is less critical than reasoning depth and tool reliability. **3. Document & Image Analysis with Long Context:** The 256K context and strong MMMU Pro (76.9%) and MATH-Vision (85.6%) scores make it suitable for processing long technical documents, PDFs, or images with complex layouts (charts, equations). It outperforms Gemma 3 27B significantly on visual tasks. **4. Fine-Tuning Base for Specialized Domains:** As a dense model with strong base performance, it's well-suited for domain-specific fine-tuning (e.g., legal, medical, scientific). Its Apache 2.0 license enables full customization and commercial deployment. Choose it over the 26B MoE for fine-tuning due to simpler gradient flow.