openbmb

/

MiniCPM-o-4_5-gguf

like 89

Follow

OpenBMB 2.87k

• LM Studio appears to use a few layers, not one single flag:

1. Catalog/model definition (model.yaml)

• metadataOverrides.reasoning: true marks it as a reasoning-capable model in LM Studio metadata/catalog.
• customFields (for example enableThinking) can define a user-facing toggle and map it to a Jinja variable in the
  prompt template.
• Source: https://lmstudio.ai/docs/app/modelyaml

2. Per-model parsing config for thought blocks

• Many LM Studio model entries include Reasoning Section Parsing with startString/endString (for example
  [THINK]...[/THINK] or seed:think...).
• That controls how LM Studio recognizes/hides reasoning sections, not whether the base model “is reasoning”
  intrinsically.
• Examples:
  • https://www.lmstudio.ai/models/mistralai/ministral-3-8b-reasoning
  • https://www.lmstudio.ai/models/bytedance/seed-oss-36b

3. Prompt-template metadata/autodetection

• LM Studio says it auto-configures prompt template from model-file metadata when available; bad/missing template
  metadata can cause behavior mismatches.
• Source: https://lmstudio.ai/docs/configuration/prompt-template