DeepSeek: DeepSeek V3.1

DeepSeek V3.1 is DeepSeek’s hybrid text model: the same named edition supports thinking and non-thinking responses through different chat-template settings. The producer highlights improved tool use, agent tasks and reasoning efficiency over its earlier model. It accepts written instructions and text context and generates an answer, code or tool-call content. The original V3.1 identity is separate from Terminus, V3.2 and the V4 family.

The model card lists 671 billion total parameters, 37 billion active parameters and a 128K context for both its base and post-trained editions. V3.1 builds on the original V3 base with extended training on long documents before post-training. Its mixture-of-experts and latent-attention structure follows the V3 architecture. These design facts explain the long-context text profile without turning the context figure into a separate maximum-output specification.

Non-thinking mode prepares a direct response, while thinking mode starts a reasoning phase before the answer. The producer template uses a closing thinking marker for direct replies and an opening thinking marker for the current reasoning turn. In a multi-turn conversation, earlier reasoning content is dropped while final replies and the closing marker remain in the context. This is the original V3.1 history rule, distinct from the later V3.2 rule that preserves reasoning across tool-related messages.

Standard tool calling is documented in non-thinking mode. The application supplies function names, descriptions and parameter schemas; the model returns a selected tool name and JSON arguments in the producer’s tool-call representation. The application executes the operation and returns its result. Give tools specific purposes and keep the returned observation linked to the call, so the next model step has evidence rather than an assumption that an action occurred.

Thinking-mode search uses a separate specialised search-agent format. DeepSeek documents a user-provided search tool for questions requiring external or current information, with repeated calls and observations over several turns. Search is therefore part of the surrounding workflow, not an assertion that a text response independently browsed the web. Define the question, the source criteria and the final answer you need before the search task starts.

For coding, supply the relevant fragment or files, observed behaviour, expected behaviour and constraints to preserve. An illustrative brief is: “Identify the cause of this failure for the supplied input, propose a focused correction and explain the condition that the change handles.” For a document task, label source sections and ask for a comparison against explicit requirements. Require the final answer to distinguish the source-supported conclusion from a proposed next action.

The producer trains V3.1 using the UE8M0 FP8 scale format for weights and activations. Its local recommendations also identify precision-sensitive parameters that must be handled correctly. These are characteristics of the published producer reference, not a claim about every application’s selected precision or implementation. A derivative, a quantized variant and the later Terminus checkpoint remain separate choices from the named original V3.1 edition.

Prepare longer work with a clear question, source labels and a defined final artifact. Budget conversation, source text, reasoning and output within the documented context and keep only relevant evidence as the task grows. For an agent, specify the allowed actions and the criterion that marks the result complete. The useful deliverable is a reviewable textual answer, code proposal or source-backed analysis tied to the input and the observations the application actually returned.