Google: Gemma 4 26B A4B (free)

Gemma 4 26B A4B IT is Google DeepMind’s instruction-tuned mixture-of-experts edition of the open Gemma 4 family. The official architecture table lists 25.2 billion total parameters and 3.8 billion active parameters; A4B names the rounded active-parameter scale. Its purpose covers reasoning, coding, agent workflows and multimodal understanding. It is separate from the pretrained edition and from the dense 31B model.

The mixture-of-experts design activates a subset of its parameters for each inference step: eight active experts from 128, plus one shared expert, in the published architecture table. Google positions this edition around inference efficiency relative to the dense model. It combines a 30-layer architecture and a 1,024-token sliding window with global attention, a 256K context and a visual encoder. These architecture figures describe the producer model, rather than a guaranteed latency or hardware configuration for a particular application.

Its 256K-token context supports a task that combines substantial text with visual evidence. Images and text can be interleaved in one prompt; the manufacturer recommends placing image content before the related text for best results. Visual tasks include object detection, document-page interpretation, screen and interface understanding, charts, multilingual OCR, handwriting and pointing. The output is text: extracted content, an explanation, code or a proposed tool action.

Configurable thinking lets the model reason before answering or give a direct answer. Gemma 4 adds a native system role alongside user and assistant roles. Google’s prompt format enables thinking with the <|think|> control token in the system prompt and separates the thought channel from the final answer; its documented processor can handle that template. For a multi-turn conversation, keep previous final answers in the history and exclude prior thoughts except in tool-call turns, where the manufacturer says they should be preserved.

Native Function Calling supports structured tool use, and Google’s launch description also documents structured JSON output. Define the task, the available function and the expected result; the application performs the requested operation and returns its output for the next step. A coding task can ask for generation, completion or correction, while a visual assistant can use the returned text to explain a chart or screen. A function-call instruction is distinct from a connected action that has actually executed.

Variable image resolution is controlled by the visual token budget. The documented options are 70, 140, 280, 560 and 1,120 tokens. Lower budgets suit classification or captions; higher budgets preserve more detail for OCR, document parsing and small text, with additional computation. Decide whether you need the overall scene or exact lettering before preparing the image. A clear crop of a label can be more useful to an extraction task than an unqualified request to inspect every detail of a full page.

Google recommends temperature 1.0, top-p 0.95 and top-k 64 as the family’s sampling configuration. The model card distinguishes training in more than 140 languages from out-of-the-box support in more than 35 languages. State the answer language and retain domain names, field names or formulas explicitly in the brief. The 256K figure is a context specification; the cited model card does not establish a separate universal maximum-output figure for this edition.

For a document-and-code assignment, supply the relevant page image, the written requirement and the source fragment: “Read the labels, explain how this screen differs from the specification and propose a focused correction to the provided code.” For a chart, ask the answer to separate visible values from an inferred trend. Gemma’s model card notes limits involving ambiguous language, very complex tasks and incorrect or outdated factual statements, so attach the authoritative facts that the requested result must use rather than expecting training knowledge to supply them.

Through Ailoot, select this exact model ID in /v1/chat/completions and supply the task in messages. Each message carries a role and its content; keep the relevant instruction and source material together. Collect the complete JSON response or all SSE segments when stream is enabled, and retain usage with the written result.