DeepSeek: DeepSeek V3.2

DeepSeek V3.2 is a text generation model from DeepSeek, released by the provider on 1 December 2025. It sits in the text category and is built around a design goal of combining computational efficiency with strong reasoning and agentic tool use. The provider describes it as introducing DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism. If you have used other DeepSeek models, this one continues that line rather than switching to a different family.

The model takes text in and returns text out. It has a context window of 163,840 tokens and a maximum output of 147,456 tokens, so long inputs and long generations are both within range. Reasoning behaviour is exposed through the include_reasoning and reasoning parameters, which means you can ask for reasoning output rather than treating the model as a plain completion endpoint. Structured outputs and response_format are also supported, so JSON-style responses can be requested directly.

Those capabilities point to a few practical uses. Long document work fits the context window: summarising, extracting, or rewriting material that would not fit in smaller models. Agent and tool-calling workflows fit the tool support, since the model accepts tools and tool_choice and can be driven in a loop where it decides which function to call. Reasoning-heavy tasks such as multi-step analysis, planning, or debugging logic benefit from the reasoning parameters. Classification, extraction, and generation tasks that need machine-readable output are served by structured_outputs and response_format. Because the model is text only, anything involving images or audio needs a different model.

You call it through api14's OpenAI-compatible API using the model id "deepseek/deepseek-v3.2". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p. If you already have an OpenAI client, the same request shape applies, with the model field set to that id and the parameters above available as needed.

Two limits are worth noting before you commit. Input and output are text only, so multimodal work is out of scope. The maximum output of 147,456 tokens is large but still finite, and very long generations may need truncation or chunking. For tasks that need image or audio understanding, look at a model in those categories instead.