DeepSeek: DeepSeek V3.1

DeepSeek V3.1 is a text model from DeepSeek, released by the provider on 21 August 2025. It is a large hybrid reasoning model with 671B parameters and 37B active, and it extends the DeepSeek-V3 base with a two-phase long-context design. Its knowledge cutoff is 31 March 2025. The model sits in the text category, so it is aimed at language tasks rather than image or audio work.

The model takes text in and returns text out. It supports both thinking and non-thinking modes, selected through prompt templates, and the reasoning parameter plus include_reasoning let you control and surface that behaviour. The context window is 163,840 tokens, with a maximum output of 32,768 tokens. Supported parameters include temperature, top_p, top_k, min_p, frequency_penalty, presence_penalty, repetition_penalty, logit_bias, logprobs, top_logprobs, seed, stop, max_tokens, response_format, structured_outputs, tool_choice and tools.

Those capabilities point to a few practical uses. Long documents fit comfortably inside the context window, so summarisation, question answering over large inputs and multi-document analysis are reasonable fits. Tool calling and structured outputs suit agents, function calling and pipelines that need machine-readable responses. The reasoning mode helps on tasks that benefit from step-by-step thinking, while the non-thinking mode keeps latency and output shorter for straightforward generation. The seed parameter supports reproducible sampling, and logprobs support classification or confidence checks.

You call the model through api14's OpenAI-compatible API using the model id "deepseek/deepseek-chat-v3.1". Because the interface follows the OpenAI chat completions shape, you send a messages array and set the model field to that id. The request parameters listed above are supported, so you can pass temperature, top_p, top_k, min_p, max_tokens, stop, seed, response_format, structured_outputs, tools, tool_choice, logit_bias, logprobs, top_logprobs, frequency_penalty, presence_penalty, repetition_penalty, reasoning and include_reasoning as needed. The page's live price table covers cost.

A few limits are worth noting. Inputs and outputs are text only, so image, audio or video tasks need a different model. The knowledge cutoff of 31 March 2025 means later events are outside its training data. Output is capped at 32,768 tokens, and the total context at 163,840 tokens, so very large jobs need chunking. If you do not need reasoning or tool calling, a smaller text model may be simpler and faster.