Gemma 4 26B A4B IT (DeepInfra)

Gemma 4 26B A4B IT is a mixture-of-experts instruction-tuned model in the Gemma 4 family, engineered to activate roughly four billion parameters per token while keeping the full 26B-parameter expert pool available for routing. That sparsity design, paired with fp8 quantization on the DeepInfra serving stack, makes the model attractive for cost-sensitive workloads that still want the depth of a larger backbone. Open weights under the Gemma license give teams room to inspect or fine-tune, while the native reasoning mode lets the model step through multi-part problems before committing to a final answer, and function calling plus structured JSON output make it straightforward to wire into agent pipelines.

On DeepInfra the deployment pairs a 262,144-token context window with a 32,768-token output cap, suiting long-document summarization, codebase analysis, and retrieval-heavy chat agents that need substantial headroom for both prompts and completions. The instruction-tuning pass targets reliable behavior on tool-using and JSON-formatted tasks, and both independent catalogs confirm function calling via an OpenAI-compatible chat completions endpoint. In practice this variant is a good fit for builders who want an open, efficient MoE model with reasoning and agent tooling on a hosted endpoint, rather than teams that need on-device inference or non-text inputs.