Google: Gemma 4 26B A4B

Google Gemma 4 26B A4B IT is an instruction-tuned model from Google DeepMind, released on 3 April 2026. It belongs to the Gemma family and uses a Mixture-of-Experts (MoE) architecture: the model holds 25.2B total parameters, but only 3.8B activate per token during inference. That design is what the "A4B" in the name refers to, and Google positions it as delivering quality close to a 31B model while keeping the active compute per token much smaller.

The model is text-in, text-out at its core, but it accepts more than text. Input modalities are image, text and video, so you can send screenshots, photos or video frames alongside a prompt. Output is text only. The context window is 262,144 tokens, and the maximum output is 235,929 tokens, which is unusually large and makes long-form generation and whole-document processing practical rather than something you have to chunk aggressively.

Those capabilities point to a few natural use cases. Long-context work such as summarising or querying large codebases, contracts, research papers or support ticket histories fits the 262k window. Because images and video are accepted as input, you can build workflows around visual question answering, extracting structured data from screenshots, or describing and reasoning about video content. Instruction tuning plus tool calling makes it suitable for agent-style pipelines, and structured outputs help when you need JSON or schema-conformant responses rather than free text.

You call it through api14's OpenAI-compatible API using the model id "google/gemma-4-26b-a4b-it". Because the interface follows the OpenAI chat completions shape, existing SDKs and clients generally work with only a base URL and key change. The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. Reasoning and include_reasoning let you control whether reasoning content is surfaced, and tools with tool_choice cover function calling.

On limits, the main thing to plan around is that output is text only, so any task requiring generated images, audio or video needs a different model. The large context and output ceilings are generous, but they do not remove the need to manage token budgets in your own application. If your workload is short, simple text generation, a smaller and faster model may be a better fit; Gemma 4 26B A4B IT makes most sense when you actually need the long context, multimodal input or tool use.