Google: Gemma 3 27B

Google Gemma 3 27B is an open model from Google, distributed under the Gemma family and served on api14 under the model id google/gemma-3-27b-it. It sits in the text category, and the provider release date is 12 March 2025. The knowledge cutoff is 31 August 2024, so anything after that date is outside what the model saw during training.

The model accepts text and image input and returns text only, which makes it a vision-language model rather than a pure text one. Its context window is 131,072 tokens, and the maximum output is 117,964 tokens, so long inputs and long generations are both viable within the same request. Google describes Gemma 3 as supporting vision-language input, handling context windows up to 128k tokens, understanding over 140 languages, and offering improved math, reasoning and chat behaviour compared with earlier Gemma releases.

Those capabilities point to a few practical uses. Document and screenshot understanding is a natural fit, since you can pass an image alongside a text prompt and get a written answer. Long-context work such as summarising large files, reviewing lengthy threads, or answering questions over a big body of text is possible without chunking in many cases. Multilingual applications benefit from the reported language coverage. Structured extraction and chat assistants are also reasonable, given the parameter list below. Reasoning and math tasks are supported, though the facts here do not include benchmark numbers, so evaluate on your own data.

You call it through api14's OpenAI-compatible API using the model id google/gemma-3-27b-it, so existing OpenAI SDKs and clients work with a base URL change. The supported request parameters are frequency_penalty, logit_bias, logprobs, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. Tool calling and structured outputs are both available, which covers function calling and JSON-shaped responses. Use seed for reproducibility, and logprobs with top_logprobs if you need token-level confidence.

A few limits are worth noting. Output is text only, so the model cannot generate images even though it reads them. The knowledge cutoff of 31 August 2024 means recent events require retrieval or prompt context. If your workload is text only and does not need vision, a smaller text model may be cheaper and faster; if you need image generation or audio, look elsewhere.