# Gemma 3 27B ![Organization](https://img.shields.io/badge/Org-Google-4285F4) ![Parameters](https://img.shields.io/badge/Params-27B-1f6feb) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Gemma%20ToU-d29922) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Mar%202025-555) > **Quick answer:** Gemma 3 27B is Google's March 2025 open-weight multimodal model, scoring 87.8% on HumanEval, 67.5% on MMLU-Pro, 89.0% on MATH, and 26.0% on BigCodeBench. With a 128K context window and Gemma Terms of Use, it is the largest and most capable model in the Gemma 3 family. ## At a Glance **Where Gemma 3 27B leads** - 89.0% MATH — strong competition mathematics performance - 87.8% HumanEval — excellent function-level code completion - 67.5% MMLU-Pro — competitive academic knowledge at 27B scale - Multimodal: text and image inputs - 128K context window - Open weights available on Hugging Face **Where it lags** - Gemma Terms of Use (not Apache 2.0 or MIT) — some commercial restrictions - 26.0% BigCodeBench — moderate complex coding - Text+vision only; no audio - Superseded by Gemma 4 family for new deployments **Best for:** Open-weight multimodal tasks at 27B scale; teams within the Google/TPU ecosystem; high-quality math and code completion tasks. ## What Gemma 3 27B Is Gemma 3 27B is the largest model in Google's Gemma 3 open-model family, released March 12, 2025. It introduces multimodal capability (text + image) to the Gemma 3 line, targeting vision-language tasks at the 27B scale that was previously only available in much larger proprietary models. The model's 89.0% MATH score is notable — placing it above many models 2-3× its parameter count on competition mathematics. Combined with 87.8% HumanEval, Gemma 3 27B is a strong choice for mathematical reasoning and code completion tasks in the open-weight 27B tier. Google deploys Gemma 3 models with JAX/TPU training, and the models are optimised for Google's infrastructure. For teams building on GCP or using Google Cloud AI Platform, Gemma 3 27B integrates well into the Google AI ecosystem. ## Specifications | Field | Value | |---|---| | Organization | Google | | Parameters | 27B (dense) | | Context window | 128,000 tokens | | License | Gemma Terms of Use | | HuggingFace | [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it) | | Release date | March 12, 2025 | | Knowledge cutoff | September 2024 | | Modality | Text + Vision (multimodal) | ## Pricing Gemma 3 27B is available as open weights under the Gemma Terms of Use. Available via Google Cloud Vertex AI and third-party providers. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [HumanEval](/benchmarks/humaneval) | 87.8% | Benchgen evaluation | 2025-07 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 67.5% | Benchgen evaluation | 2025-07 | | [MATH](/benchmarks/math) | 89.0% | Benchgen evaluation | 2025-07 | | [BigCodeBench](/benchmarks/bigcodebench) | 26.0% | Benchgen evaluation | 2025-07 | ## Gemma 3 27B vs Alternatives | Model | HumanEval | MMLU-Pro | MATH | License | |---|---|---|---|---| | **Gemma 3 27B** | 87.8% | 67.5% | 89.0% | Gemma ToU | | [Gemma 3 12B](/models/google/gemma-3-12b) | 85.4% | 60.6% | — | Gemma ToU | | [Phi-4](/models/microsoft/phi-4) | 82.6% | 70.4% | — | Apache 2.0 | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | — | 80.5% | — | Llama 4 | Gemma 3 27B vs Phi-4: higher HumanEval and MATH at 27B vs 14B — Phi-4 has higher MMLU-Pro (70.4% vs 67.5%) at smaller size. Gemma 3 27B is the preferred choice when multimodal capability is needed. ## Related Models & Benchmarks - Compare: [Gemma 3 12B](/models/google/gemma-3-12b) · [Phi-4](/models/microsoft/phi-4) · [Llama 3.1 70B Instruct](/models/meta/llama-3-1-70b-instruct) - Benchmarks: [HumanEval](/benchmarks/humaneval) · [MMLU-Pro](/benchmarks/mmlu-pro) · [MATH](/benchmarks/math) ## Frequently Asked Questions

What is Gemma 3 27B?

Gemma 3 27B is Google's March 2025 open-weight multimodal 27B model, scoring 87.8% HumanEval, 67.5% MMLU-Pro, and 89.0% MATH with a 128K context window.

Is Gemma 3 27B open source?

The weights are open under the Gemma Terms of Use on Hugging Face. This license permits commercial use with some conditions — review the terms for your specific use case.

What is Gemma 3 27B's context window?

Gemma 3 27B supports a 128,000-token context window.

Does Gemma 3 27B support images?

Yes. Gemma 3 27B is a multimodal model supporting both text and image inputs.

--- *Specs from Google's official Gemma 3 release (March 2025) and Benchgen evaluations. Last updated 2026-07-24.*

Gemma 3 27B — Evaluations

Evaluations