Activefastllm Open Source

Gemma 3 27B

by Google· Released March 2025· Cutoff February 2025

Gemma 3 27B is Google's latest open model in the Gemma family, offering strong performance for its size with a 128K context window and multilingual support. It is designed for efficient deployment on consumer hardware while delivering competitive results on reasoning, coding, and instruction-following tasks.

Official Site API Docs 🤗 Hugging Face 📄 Research Paper

Input cost

Free (open source)

Output cost

Free (open source)

Context window

128K tokens

Max output

8192 tokens

Modalities

text

Parameters

27B

License

Gemma License

Capabilities

Text GenerationCode GenerationMultilingual SupportFunction CallingStructured OutputInstruction FollowingReasoning

Best For

Efficient, high-quality text generation and coding tasks on consumer hardware with a large context window.

Strengths

Strong performance for its size, competitive with larger models
128K context window enables long-document processing
Open source and free to use, deployable on consumer GPUs
Multilingual capabilities with support for over 140 languages

Limitations

Not multimodal (text-only input)
Smaller parameter count may limit performance on complex reasoning compared to frontier models
No native vision or audio support
Requires significant GPU memory for full-precision inference

Use Cases

Chatbots and virtual assistants

Code generation and debugging

Document summarization and analysis

Multilingual translation and content creation

Educational tutoring and Q&A

Data extraction and structured output generation

Long-form content writing

Improvements Over Previous Model

Context window increased from 8K to 128K tokens compared to Gemma 2
Added multilingual support for over 140 languages (Gemma 2 was primarily English)
Improved reasoning and coding benchmarks (e.g., HumanEval 75.2% vs 60.4% in Gemma 2 27B)
Introduced sliding window attention for efficient long-context processing
Better instruction following and safety alignment via RLHF
Support for function calling and structured output (JSON mode)

Back to all models

Improvements Over Previous Model

Context window increased from 8K to 128K tokens compared to Gemma 2

Added multilingual support for over 140 languages (Gemma 2 was primarily English)

Improved reasoning and coding benchmarks (e.g., HumanEval 75.2% vs 60.4% in Gemma 2 27B)

Introduced sliding window attention for efficient long-context processing

Better instruction following and safety alignment via RLHF

Support for function calling and structured output (JSON mode)