Google: Gemma 4 26B A4B (free)

Gemma 4 26B A4B IT is an instruction-tuned Mixture-of-Experts model from Google, published under the Gemma family from Google DeepMind. It has 25.2B total parameters, but only 3.8B activate per token during inference, which is the design choice behind the A4B name. It is a text model in api14's catalogue, and the provider release date is 2026-04-03. The version served here is the free variant, identified as google/gemma-4-26b-a4b-it:free.

On the input side, the model accepts text, images and video, and it produces text only. The context window is 262144 tokens, and the maximum output is 32768 tokens. Because it is instruction tuned, it is meant to follow prompts directly rather than continue raw text. The mixture-of-experts layout means you get a relatively large total parameter count while only a small fraction of weights is active per token, which is relevant if you care about throughput or serving efficiency rather than raw parameter count alone.

Those capabilities point to a few practical patterns. Long documents, large codebases or extended transcripts fit comfortably inside the 262144 token context, so summarisation, question answering over a whole repository, and multi-document analysis are reasonable fits. Mixed-media inputs make it usable for tasks that combine screenshots, diagrams or video frames with written instructions, such as extracting structured details from a recorded walkthrough or describing what happens in a clip alongside a text prompt. Tool calling and structured output support make it a candidate for agent-style workflows and for pipelines where the response must conform to a schema.

You call it through api14's OpenAI-compatible API using the model id "google/gemma-4-26b-a4b-it:free". The supported request parameters are include_reasoning, max_tokens, reasoning, response_format, seed, temperature, tool_choice, tools and top_p. In practice that means you can set sampling behaviour with temperature and top_p, make runs reproducible with seed, cap the reply length with max_tokens, request JSON or another structured shape with response_format, and pass tool definitions with tools and tool_choice. The reasoning and include_reasoning parameters control whether reasoning output is requested and whether it is returned.

Keep the 32768 token output ceiling in mind when planning long generations, and note that the model returns text only, so any image or video handling happens on the input side. If your task is purely text and does not benefit from the long context or the multimodal inputs, a smaller text-only model may be simpler to work with.