[tc-mb]

Overview

MiniCPM-OCR 1.0 is the MiniCPM-V 4.6 stack with a different model_type, so
this only registers it: a new PROJECTOR_TYPE_MINICPM_OCR_1_0 reusing
clip_graph_minicpmv4_6 and the 4.6 conversion path.

Additional information

The model writes its transcription inside the thinking block, so run with
--reasoning off (or reasoning_effort: "none" per request) to get it in
content. With the default auto the reply comes back as empty content and
the text lands in reasoning_content.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — AI tools were used for part of the code changes.

[ngxson]

chat_template_kwargs={"enable_thinking": false} is needed for plain
transcription, otherwise the template emits <think>.

that is --reasoning off

which AI agent you are using? seems like it didn't discover the code base deeply

[tc-mb]

chat_template_kwargs={"enable_thinking": false} is needed for plain
transcription, otherwise the template emits <think>.

that is --reasoning off

which AI agent you are using? seems like it didn't discover the code base deeply

You are correct; I seem to have conflated the AI-related terminology with the specific usage within vLLM. I have updated.