Back to Models
DeepMindOpen Source

Gemma 4 26B A4B(混合专家模型)

Gemma 4 26B A4B is a foundation model developed by DeepMind, employing a Mixture of Experts (MoE) architecture. With approximately 25.2B parameters, it is optimized for chat-format interactions.

Parameters

25.2B

Context Window

256K

License

Apache 2.0

Release Date

2026-04

API Pricing

API pricing for this model is not yet available

Strengths

  • Efficient reasoning via MoE
  • 256K long context window
  • Open Apache 2.0 license

Weaknesses

  • Lack of specialization for specific tasks
  • Performance gap compared to larger models
  • Potential fluctuations in operational costs

Use Cases

  • Advanced chatbot development
  • Long-form document analysis
  • Building open-source AI

Deep Analysis

Arena Elo (Text Overall)

1438

From BenchLM (7,777 votes)

Arena Elo (Coding)

1481

Strong coding-specific performance

GPQA Diamond (Reasoning)

82.3%

vs Gemma 4 31B: 84.3%

AIME 2026 (Math)

88.3%

High math-reasoning capability

Active Parameters

~3.8B

Out of 25.2B total (MoE)

Input/Output Price

$0.13 / $0.40 per 1M tokens

Blended ~$0.20/M

Strengths

  • Exceptional efficiency: MoE architecture with only ~3.8B active parameters delivers performance rivaling much larger dense models.
  • Strong mathematical and coding reasoning, with high AIME and LiveCodeBench scores.
  • Apache 2.0 license simplifies commercial adoption and deployment compared to previous Gemma versions.

Weaknesses

  • Overall Arena Elo (1438) lags behind top proprietary models (e.g., Gemini 3.1 Pro) and even its dense 31B sibling.
  • Agentic performance is a noted weakness, with low scores on benchmarks like Terminal-Bench (13.6) and HLE (8.7).
  • Lacks native audio support found in smaller Gemma 4 E2B/E4B models, limiting some multimodal use cases.

Competitor Comparison

ModelArenaSWEGPQAPrice
Gemini 3.1 Pro (Google)~1480+N/A~92%+Proprietary/Subscription
Gemma 4 31B Dense (Google)145280.0%*84.3%$0.13/$0.40 per 1M tokens
Llama 4 (Meta)~1460*N/A~89%*Open Weight

Gemma 4 26B A4B is Google DeepMind's efficiency-focused open-weight model released in April 2026. As a Mixture-of-Experts (MoE) model, its defining characteristic is achieving near-frontier performance with minimal active parameters (~3.8B), making it highly cost-effective for inference while handling substantial 256K context windows. This positions it as a compelling "sweet spot" for developers and researchers who need strong capabilities—particularly in reasoning, coding, and multimodal tasks—without the computational demands of its larger 31B dense sibling or the resource requirements of proprietary frontier models.

The model excels in structured reasoning tasks, demonstrated by top-tier scores on benchmarks like AIME (math) and LiveCodeBench (coding). Its release under the permissive Apache 2.0 license marks a significant shift from earlier Gemma models, greatly simplifying commercial and private deployment. While it may not surpass the absolute best proprietary models in every benchmark, its combination of strong performance, architectural efficiency, and open licensing makes it a highly practical choice for a wide range of applications, from on-device assistive agents to large-scale enterprise pipelines.

Analysis generated: 2026-05-23