Skip to main content

DeepSeek V3.2 / DSML

DeepSeek V3.2 is the reference point for DSML-based prompt and parsing behavior.

Why DeepSeek V3.2 Is Special​

DeepSeek V3.2 uses a DSML-oriented flow instead of the generic tokenizer chat template and generic output parsing path.

That means the runtime needs to know:

  • when to use DSML prompt encoding
  • when to parse DSML-style reasoning and tool-call output
  • when a model should be treated as a DeepSeek V3.2 family member even if the tokenizer metadata alone is not enough

Current Capability Handling​

Skulk supports DeepSeek V3.2 through resolved family defaults:

  • DSML prompt renderer
  • DeepSeek V3.2 output parser
  • DSML tool-call format
  • thinking-toggle aware family behavior

Built-In Card Status​

The validated built-in mlx-community/DeepSeek-V3.2-{4bit,8bit} cards now declare the same DSML contract that the runtime already used through resolved family defaults:

  • toggleable reasoning metadata
  • DSML tool-call format
  • DSML prompt rendering
  • DeepSeek V3.2 output parsing

The resolved capability layer still matters for compatibility and future aliases, but the trusted V3.2 quantizations no longer rely on hidden family inference alone.

Why This Matters​

DeepSeek V3.2 demonstrates that the capability system can represent a model family whose key differentiation is not vision, but a custom prompt/parser contract.