b10254

ggml-org/llama.cppb10254Aug 4, 2026by github-actions[bot]same day as b10253

What changed since b10253

medium risk

Adds a DeepSeek V4 Flash 0731 chat template with structured-output prompts and reworked reasoning-effort handling.

Breaking changes

  • DeepSeek V4 chat history now drops prior thinking by default unless preserve_reasoning is requested or tools are present, changing prompt content for existing V4 users.

New features

  • New chat template for DeepSeek V4 Flash 0731 with updated high/max reasoning-effort mapping.
  • Structured output response-format instructions added to the DeepSeek V4 templates, with the schema passed into template rendering.
  • Template rendering tests now cover reasoning effort, drop_thinking, structured output prompts, preserved reasoning, continuations, and empty tool arguments.

Improvements

  • DeepSeek V4 templates aligned with the official encoders while parser behavior is kept out of this change.
  • Template selection logic for the DeepSeek V4 0731 variant corrected.
  • Minor cleanup removing an unneeded lowercase normalization step.

Bug fixes

  • DSML parser now consumes the tool-call separator instead of leaving it in the output.

Upgrade notes

No configuration or migration steps are required; the new template is selected automatically for DeepSeek V4 Flash 0731 models. If you depend on prior assistant thinking being retained in DeepSeek V4 history, set preserve_reasoning explicitly, since drop_thinking is now the default. Structured-output schemas are now injected into V4 template rendering, so verify prompts if you use response_format.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Introduces a new chat template for DeepSeek V4 Flash 0731 with support for structured output and reasoning effort mapping.

Key Highlights

  • DeepSeek V4 Flash 0731 template
  • Structured output response format instructions
  • Updated reasoning effort mapping

Release Notes (excerpt)

<details open> chat : add new template for DeepSeek V4 Flash 0731 (#26398) * common/chat: update DeepSeek V4 templates Align the DeepSeek V4 templates with the official encoders while keeping parser behavior out of this change. - Default drop_thinking for DeepSeek V4 history so prior thinking is omitted unless preserve_reasoning is requested or tools are present. - Add structured output response-format instructions to the V4 templates and pass the schema into template rendering.…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories