Google: Gemma 4 31B

Google: Gemma 4 31B is an instruction-tuned model from Google, listed on api14 under the model id google/gemma-4-31b-it. It belongs to the Gemma family from Google DeepMind, and the provider description identifies it as a 30.7B dense multimodal model. Its main category on api14 is text, and it is offered by Google as the provider. If you are choosing between models on the gateway, this one sits in the mid-size dense range rather than the very largest hosted options, which is relevant when you weigh capability against latency and throughput for your workload.

The model accepts image, text and video as input and returns text only. That combination makes it a multimodal reader rather than a generator of media: you can send it documents, screenshots, photos or video alongside a prompt, and it will answer in prose. The context window is 262144 tokens, which is large enough to hold long documents, extended conversations or sizeable media transcripts in a single request. The maximum output is 16384 tokens, so responses are bounded well below the input capacity. It also supports a configurable thinking or reasoning mode, exposed through the reasoning and include_reasoning parameters, which lets you decide whether the model spends tokens working through a problem before answering.

Those capabilities point to a fairly specific set of uses. Long-context multimodal reading is the obvious one: extracting structured information from scanned documents, PDFs rendered as images, or screenshots, and summarising or answering questions about them. Video input extends this to tasks like describing footage, pulling out events or answering questions about what happens in a clip. The reasoning mode suits multi-step analysis where a direct answer is unreliable, and the tool and function support makes the model usable as the decision-making layer in an agent that calls external services. Structured outputs and response_format let you constrain replies to JSON or a schema when the result feeds another system.

Calling it through api14 works the same way as any other model on the gateway, since the API is OpenAI-compatible. You send a chat completion request with the model field set to google/gemma-4-31b-it and your messages in the standard format, including image or video content parts where relevant. The supported parameters cover the usual sampling controls (temperature, top_p, top_k, min_p, seed, stop, frequency_penalty, presence_penalty, repetition_penalty), output controls (max_tokens, response_format, structured_outputs), inspection options (logprobs, top_logprobs, logit_bias) and tool use (tools, tool_choice). Reasoning behaviour is controlled with reasoning and include_reasoning.

One practical constraint to keep in mind is the output ceiling: 16384 tokens per response, which is modest relative to the 262144 token input window. If your task is generating very long documents in one pass, a model with a larger output allowance may fit better. Similarly, if you need audio input or any non-text output, this model does not cover it, since its input modalities are image, text and video and its output is text only. For long-context multimodal understanding, reasoning-heavy analysis and tool-driven agents, it is a reasonable fit.