Question: How can I enable DeepSeek V4 Flash 0731 Max reasoning mode?

[BeCreatedP]

Hi, thank you for adding DeepSeek V4 Flash 0731 and DSpark support in oMLX v0.5.4rc2.

I have a question about the reasoning-effort setting for this model.

In the model settings, the built-in reasoning_effort selector currently provides:

• low
• medium
• high

DeepSeek V4 Flash 0731 also appears to have a Max reasoning mode in its official inference/encoding implementation.

What is the correct way to enable the official Max reasoning mode when running the model through oMLX?

For example, should I:

1. add a custom chat-template argument such as:

   {
     "enable_thinking": true,
     "reasoning_effort": "max"
   }

2. use a particular system-prompt prefix;

3. modify the model’s tokenizer/chat-template configuration; or

4. use another supported setting?

I am not sure whether the current high option already maps to DeepSeek’s highest local reasoning mode, or whether max needs to be selected separately.

Environment:

• oMLX: v0.5.4rc2
• Model: DeepSeek-V4-Flash-0731 with embedded DSpark weights
• Hardware: Mac Studio M3 Ultra, 512 GB unified memory
• Thinking mode: enabled
• Lightning MTP / DSpark: enabled

I would appreciate clarification on the recommended configuration. Thank you.

[jundot]

Thanks for the detailed report. high does not map to DeepSeek V4's Max mode; max is a separate reasoning_effort value in the official encoder, and it applies the Max prefix when thinking mode is enabled. I've added max to both the web admin and macOS model settings selectors. Until this is included in a release, setting chat_template_kwargs to {"enable_thinking": true, "reasoning_effort": "max"} enables Max mode without a custom system prompt or tokenizer change.