AI Radar

Gemma 2

Available · Language · Milestone

Gemma 2 is the second generation of Google's Gemma open-weight language models, released in 9B and 27B sizes with weights on Kaggle and Hugging Face, followed by a 2B size in July 2024. Google says it uses a redesigned architecture aimed at performance and inference efficiency, and it trained the 2B and 9B models with knowledge distillation. [1] [2] [3] Primary source

Timeline of Gemma 2 →

Claims and evidence

  • Announced Primary source[4]Next generation of open models [5]Additions to the Gemma family
  • Released 9B and 27B sizes; the 2B size followed on 2024-07-31 (see variants). Primary source[1] [6]June 27, 2024
  • Status AvailableWeights still downloadable from Google's Hugging Face organisation (after accepting the Gemma terms) when checked on 2026-10-01. Primary source[7] [8] [9]
  • Successor of Gemma 1.1 EditorialEditorial link along the Gemma line: the announcement compares Gemma 2 with the first generation without naming a version; Gemma 1.1 was the latest first-generation release, and the technical report extends the Gemma 1.1 post-training data and compares with Gemma 1.1. Primary source[1]second paragraph [3]sections 2 (Model Architecture) and 4 (Post-Training)
  • Change · Architecture Alternates local sliding-window and global attention layers and adds grouped-query attention and logit soft-capping, with deeper networks.Compared with Gemma 1.1 Primary source[3]section 2, Model Architecture
  • Change · Training data The 2B and 9B models are trained with knowledge distillation from a larger teacher model instead of next-token prediction; the 27B model is trained from scratch.Compared with Gemma 1.1 Primary source[3]
  • Change · Size Adds 9B and 27B sizes; the small model that followed in July 2024 has about 2.6B parameters.Compared with Gemma 1.1 Primary source[1] [3]Table 2
  • Change · Other Post-training extends the Gemma 1.1 data with more internal and public data and uses a similar RLHF algorithm with a much larger reward model.Compared with Gemma 1.1 Primary source[3]section 4, Post-Training
  • Input text Primary source[10]Inputs and outputs
  • Output text Primary source[10]Inputs and outputs
  • Open weights YesReleased under Google's Gemma terms of use, which must be accepted on Hugging Face before download. Primary source[1]Getting started [8]
  • Context window 8192 tokens Primary source[3]section 2, Model Architecture
  • Access Open-weights downloadWeights on Kaggle and Hugging Face; Google also offered the 27B model to try in Google AI Studio and announced Vertex AI deployment. Primary source[1]Getting started

Lineage

Predecessors

Successors

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

Gemma 2 2B

  • Released Primary source[2] [6]July 31, 2024
  • Variant Inline variant in this record.Pre-trained and instruction-tuned checkpoints. Google says the 2B model learns from larger models through distillation. Primary source[2] [7]

Also known as: gemma-2-2b, gemma-2-2b-it

Differs in:

  • Parameters: c. 2.6B Evidence not assessed [1]Projects built with Gemma [3]Table 2

Gemma 2 9B

Same dates as Gemma 2.

  • Variant Inline variant in this record.Pre-trained and instruction-tuned checkpoints. Primary source[1] [8]

Also known as: gemma-2-9b, gemma-2-9b-it

Differs in:

  • Parameters: c. 9B Evidence not assessed [1]second paragraph [3]Table 2

Gemma 2 27B

Same dates as Gemma 2.

  • Variant Inline variant in this record.Pre-trained and instruction-tuned checkpoints. Primary source[1] [9]

Also known as: gemma-2-27b, gemma-2-27b-it

Differs in:

  • Parameters: c. 27B Evidence not assessed [1]second paragraph [3]Table 2

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Google on AI Radar →