Amazon Bedrock adds Gemma 4 models with 256K context and built-in reasoning mode

Amazon Web Services today announced availability of Google DeepMind's Gemma 4 family on Amazon Bedrock. The open-weight models include three instruction-tuned variants spanning 2.3B to 30.7B parameters, with 256K context windows, multimodal input support, and built-in reasoning mode.

Three model variants

The Gemma 4 family includes:

• Gemma 4 31B: Dense architecture with 30.7B parameters, 256K context window
• Gemma 4 26B-A4B: Mixture-of-experts design with 25.2B total parameters but only 3.8B active per token, 256K context window
• Gemma 4 E2B: Compact model with 5.1B total parameters (2.3B effective), 128K context window

All three variants support text and image input, native function calling, and over 35 languages. According to AWS, independent benchmarks from Artificial Analysis report an Intelligence Index of 39 for Gemma 4 31B, compared to a median of 15 in the 4B-40B open-weights class.

Technical architecture

The models use hybrid attention that interleaves local and global attention to maintain long context support while reducing memory footprint. The 26B-A4B variant activates only 3.8B parameters per token despite having 25.2B total, delivering what AWS describes as "4B-class cost and latency with the knowledge capacity of a larger model."

The E2B variant uses Per-Layer Embeddings (PLE) to keep its effective parameter count at 2.3B of 5.1B total parameters.

Built-in reasoning mode

All Gemma 4 variants include a built-in reasoning mode that, when enabled, emits the model's internal thought process before producing the final answer. AWS documentation notes that in multi-turn conversations, only final answers from previous turns should be sent back to the model, not their reasoning items, as "replaying prior reasoning back to the model can degrade its responses."

Service access

The models are accessed through Amazon Bedrock's bedrock-mantle endpoint, which uses an OpenAI-compatible API. The endpoint URL is https://bedrock-mantle.{region}.api.aws/openai/v1 and supports both Chat Completions and Responses APIs.

All three variants are available in Standard, Priority, and Flex service tiers. AWS states that prompts and completions are not used to train any models and content is not shared with third parties.

The models are released under the Apache 2.0 license, allowing independent evaluation of model architecture and training methodology.

What this means

Gemma 4's availability on Bedrock gives enterprises access to competitive open-weight models through AWS infrastructure without managing inference stacks. The MoE variant's 3.8B active parameters at 25.2B total capacity represents a meaningful efficiency gain for high-throughput workloads. The 256K context window matches or exceeds most competing models, though pricing details were not disclosed in the announcement, making direct cost comparisons premature.