Google: Gemma 3 12B

Gemma 3 12B is a text generation model from Google, released on 13 March 2025 as part of the Gemma 3 family. It is the instruction-tuned 12B variant, identified on api14 as google/gemma-3-12b-it. Gemma 3 marks Google's move to bring multimodality to the Gemma line, so this model accepts both text and images as input while producing text as output.

The model supports a context window of up to 131072 tokens and can generate up to 16384 tokens in a single response. Its knowledge cutoff is 31 August 2024. According to Google, Gemma 3 handles over 140 languages and brings improved math, reasoning, and chat behaviour compared with earlier Gemma releases. Because it accepts image input alongside text, it can work with vision-language prompts rather than text alone, though everything it returns is plain text.

Those capabilities point to a fairly broad set of practical uses. The long context window suits tasks that involve large documents, such as summarising long reports, answering questions over a body of text, or working through extended chat histories. Multimodal input makes it a reasonable fit for describing or reasoning about images, for example extracting information from screenshots, diagrams, or photos and returning a written answer. Its multilingual coverage and improved math and reasoning make it useful for translation, structured problem solving, and general assistants. Tool calling and structured outputs also allow it to sit inside pipelines that need machine-readable responses or calls to external functions.

You call the model through api14's OpenAI-compatible API using the model id google/gemma-3-12b-it, so existing OpenAI client libraries work with only the base URL and key changed. The request parameters it supports are frequency_penalty, logit_bias, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, and top_p. That mix covers sampling controls, deterministic output via seed, stop sequences, JSON or schema-constrained responses through response_format and structured_outputs, and tool use through tools and tool_choice.

One thing to keep in mind is the output ceiling: responses are capped at 16384 tokens even though the context window is much larger, so very long generations need to be split or continued across calls. The knowledge cutoff of 31 August 2024 also means the model has no awareness of events after that date unless you supply the information in the prompt. If your task is text-only and does not benefit from image input, a smaller or more specialised text model may be cheaper and faster, but for mixed text and image work within a long context, this model is a sensible default.