Google: Gemma 4 31B (free)

Gemma 4 31B Instruct is an open model from Google DeepMind, published by Google, and offered on api14 under the model id google/gemma-4-31b-it:free. It is a 30.7B dense multimodal instruction-tuned model, part of the Gemma family, and it sits in the text category on api14 even though it accepts more than text as input. The provider release date is 2026-04-02.

On the input side, the model accepts text, images and video, and it produces text only. Its context window is 262,144 tokens, with a maximum output of 32,768 tokens. It supports a configurable thinking or reasoning mode, which you can steer through the reasoning and include_reasoning parameters. The spread of inputs means you can mix prose, screenshots, photos and video frames in the same conversation without switching models.

Those capabilities point to a fairly wide set of practical uses. Long documents, codebases, transcripts and multi-hour conversation logs fit inside the 256K context, so summarisation, question answering and extraction over large inputs are reasonable fits. Because images and video are accepted, tasks like describing visual content, reading charts or screenshots, moderating user-submitted media and answering questions about recorded footage are also viable. Instruction tuning plus tool support makes it a candidate for assistants that need to call functions, return structured output, or reason step by step before answering.

Calling it through api14 works the same way as any other model on the gateway, since the API is OpenAI-compatible. You send your request to the chat completions endpoint and set the model field to google/gemma-4-31b-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 control sampling with temperature and top_p, cap the reply with max_tokens, make runs reproducible with seed, request JSON or other structured shapes with response_format, and define callable functions with tools and tool_choice. If you want the model's thinking surfaced or hidden, use reasoning and include_reasoning.

A few things are worth keeping in mind before you commit. Output is text only, so if you need generated images, audio or video, this model will not produce them. The 32,768 token output ceiling is well below the input context, so very long generations, such as entire books in one response, are not what it is built for. There is no stated pricing or rate information here, so check the live price table on the page for current terms. Within those bounds it is a straightforward choice for multimodal understanding, long-context work and tool-using assistants.