Gemma 2 27B Instruct

Gemma 2 is Google's open model family, derived from the same research lineage as Gemini. The 27B-it (instruction-tuned) checkpoint hits a sweet spot: meaningfully stronger than Gemma 2 9B, still fits comfortably on a single 48 GB workstation card or a pair of consumer 24 GB cards.

Architecture details worth knowing

Gemma 2 uses alternating local + global attention layers (4096-token local window) and logit soft-capping during training. These choices make it slightly idiosyncratic to run — older inference frameworks may need updates to support it correctly.

What it's good at

The 27B is a strong generalist for its size. English is its primary strength but it handles major European languages reasonably. It's particularly good at following style and tone instructions, making it a common choice for content-generation pipelines.

What to watch for

The 8K context window is short by 2026 standards. If you need 32K+ context, look elsewhere or use one of the Gemma 2 long-context community fine-tunes. Knowledge cutoff is mid-2024.

License

Gemma Terms of Use — permissive for most commercial and research use, with a usage-policy attachment that prohibits specific harmful applications. Google does not require attribution or fee for redistribution but does require the policy to travel with derivatives.