Gemma 2 27B

Gemma 2 27B is a dense, text-only transformer in Google's open-weights Gemma family, built around 27 billion parameters and designed for conversational and general text generation tasks. According to the apxml.com specifications page, its architecture draws on standard modern transformer practices, using RoPE positional embeddings, RMSNorm normalization, 46 layers, and grouped-query attention configured with 32 query heads against 16 key-value heads at a head dimension of 128. The model is released under the Gemma License, sits in the open-weights category, and targets developers who want a mid-to-large language model that can be self-hosted or run through third-party inference providers.

Practically, Gemma 2 27B is aimed at workloads that fit inside its 8,192-token context window, such as short-document question answering, chat, summarization, and instruction following, rather than long-context retrieval or multimodal tasks. Third-party evaluation snapshots place it in the middle of the pack on general knowledge and reasoning suites, including an MMLU score around 0.75 and competitive positions on text arena and ProLLM stack evaluations, giving it a balanced profile for a model of its size. It predates the multimodal Gemma 3 generation, which introduced vision-language input and a much larger context window, so Gemma 2 27B remains a sensible pick for teams that specifically want an open-weights text model with a modest context budget.