DeepSeek's current docs now expose paired non-streaming and streaming thinking-mode samples, giving developers a first-party baseline for `reasoning_content` replay and chunk handling

Checked on July 8, 2026: DeepSeek's official docs sitemap currently lists dedicated thinking-mode sample pages for both non-streaming and streaming chat completions, clarifying when `reasoning_content` is just captured, when it is ignored in ordinary follow-up turns, and how `delta.reasoning_content` should be accumulated separately from visible answer text.

Accepted official-source monitoring note

Today's publish-safe item stays DeepSeek-first and uses backup official documentation because the live @deepseek_ai timeline still was not safely readable in this run. DeepSeek's English homepage still anchors the official X account, but the stronger current developer signal is in the official docs sitemap and the dedicated thinking-mode sample pages it currently exposes.

What we verified on July 8, 2026

DeepSeek's official docs sitemap currently lists dedicated non-streaming and streaming thinking-mode sample pages.
The non-streaming sample shows a plain two-turn OpenAI-compatible flow: capture reasoning_content and final content, append the assistant message, then send a second user turn.
The same non-streaming page explicitly notes that earlier reasoning_content will be ignored by the API in that ordinary follow-up case, which is an important boundary for teams that over-engineer replay behavior.
The streaming sample shows delta.reasoning_content and delta.content arriving separately, and its code accumulates them into different buffers before constructing the next-turn assistant state.
This is protocol and support coverage only. It does not create any new plan card, stock promise, or pricing change in this repo.

Why this is publishable

This is a current first-party developer reference from official DeepSeek docs, not a community wrapper or second-hand screenshot.

It gives developers a more complete baseline for thinking mode than yesterday's tool-call-specific sample alone.
It is distinct from the July 6 transcript coverage because today's angle is transport and turn-shape: non-streaming state capture versus streaming chunk handling.
It answers two real support questions without speculation: when reasoning_content can be ignored in ordinary follow-up turns, and how streaming chunks should be split between reasoning and visible answer buffers.
It creates a clean source cluster for today's two new guide pages without recycling the July 5 alias-retirement warning or the older V4 preview anchor.

Rejected candidates today

Direct official X timeline as the primary source: rejected because the timeline still was not safely readable here.
The July 24 alias-retirement warning: still official, but already used in the July 5 publish and therefore too duplicative for today's headline.
The tool-call sample and transcript pages alone: still official and useful, but already represented in the July 6 publish, so weaker than the paired non-streaming plus streaming angle for today's single news slot.
The official status page: checked and healthy, but weaker than the newly emphasized sample-pair contract for today's headline.
Official GitHub and Hugging Face surfaces: checked as backup official sources, but neither exposed a clearer current DeepSeek API support signal than the sample-pair docs pages.

Editorial takeaway

The clearest current DeepSeek developer update is not another launch recap. It is that the official docs now give teams first-party implementation fixtures for both ordinary non-streaming thinking mode and streaming thinking mode. That reduces guesswork around reasoning_content, replay boundaries, and chunk accumulation, which makes it safer to build or debug real DeepSeek clients.