def get_customizable_model_schema(
        self, model: str, credentials: dict
    ) -> AIModelEntity:
    schema = {item.model: item for item in self.predefined_models()}.get(
        _base_model(model)
    )
    if schema is None:
        # Fallback for custom / non-OpenAI base models served via an
        # OpenAI-compatible endpoint (e.g. Qwen, DeepSeek, local vLLM):
        # clone a chat-capable predefined schema so the model is listable
        # and can be set as the default model.
        template = next(
            (m for m in self.predefined_models()
             if m.model_properties.get("mode") == "chat"),
            None,
        )
        if template is None:
            template = self.predefined_models()[0]
        props = dict(template.model_properties)
        mp = (credentials or {}).get("model_properties") or {}

        # Align context_size to the actual vLLM --max-model-len (65536) so Dify
        # never lets prompts exceed what the served model can accept. The
        # credential-update API does NOT persist model_properties for this
        # provider, so we hardcode the safe ceiling here as the default and
        # still honor an explicit override when one is supplied.
        VLLM_MAX_CTX = 64000
        try:
            props["context_size"] = int(mp["context_size"]) if "context_size" in mp else VLLM_MAX_CTX
        except (ValueError, TypeError):
            props["context_size"] = VLLM_MAX_CTX

        # NOTE: do NOT put max_tokens into model_properties.
        # Dify API-side ModelPropertyKey only allows {mode, context_size, ...}
        # and does NOT include MAX_TOKENS; adding it breaks Pydantic validation
        # when the API parses the schema from plugin_daemon.
        # max_tokens is already exposed as a parameter_rule by the cloned template.
        schema = template.model_copy(deep=True, update={"model_properties": props})
    return schema.model_copy(
        deep=True,
        update={
            "model": model,
            "label": I18nObject(zh_hans=model, en_us=model),
            "fetch_from": FetchFrom.CUSTOMIZABLE_MODEL,
        },
    )