Gemma 3 27B Instruct

Gemma 3 27B Instruct is an instruction-tuned, open-weight model developed by Google DeepMind. It is part of the Gemma 3 family and was released in 2025 as a lightweight, high-performance alternative to Gemini. This model supports text generation and is optimized for multi-turn dialogue, summarization, and reasoning tasks.

The model is designed for:

• Chatbots and conversational AI
• Code generation and reasoning (HumanEval: 48.8)
• Document and image summarization (text only on Fireworks)
• Question answering (MMLU: 78.6, GSM8K: 82.6)
• Multilingual tasks in over 140 languages

The model supports a maximum context length of 131.1k tokens.

Gemma 3 models perform reliably across long inputs, but no specific usable token window is provided. Performance may vary near the 131k token upper bound.

Limitations and risks include:

• Limited spatial and common-sense reasoning
• Occasional factual errors or outdated knowledge
• May reflect bias from training data despite filtering
• English-only evaluation for ethics and safety benchmarks

No, function calling is not supported for this model.

The model has 28.4 billion parameters (not a MoE architecture).

Yes, Fireworks supports LoRA fine-tuning, serverless LoRA, and full fine-tuning for Gemma 3 27B Instruct.

On-demand deployment is supported with no rate limits.

Gemma 3 27B Instruct is governed by the Gemma license, which permits commercial use.