mlx-vlm Command A Vision responses contain only "<PAD>" tokens

No matter what image or prompt I provide when running the mlx-community/command-a-vision-07-2025-4bit model, the output is the system and default preambles followed by "<PAD>" repeated a bunch of times. See example below.

Do I need to provide some additional parameters when using this model in order to get a proper response? Or perhaps this is a bug? Or a limitation of the model itself?

python -m mlx_vlm.generate --model mlx-community/command-a-vision-07-2025-4bit --max-tokens 500 --temperature 0.0 --prompt "Describe this image." --image ~/Downloads/oilwelllandscape.jpeg
<frozen runpy>:128: RuntimeWarning: 'mlx_vlm.generate' found in sys.modules after import of package 'mlx_vlm', but prior to execution of 'mlx_vlm.generate'; this may result in unpredictable behaviour
Calling `python -m mlx_vlm.generate ...` directly is deprecated. Use `mlx_vlm generate` or `python -m mlx_vlm generate` instead.
Fetching 21 files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 21/21 [00:00<00:00, 28532.68it/s]
/Users/dwohlfahrt/workspace/github/dwohlfahrt/mlx-testbed/.venv/lib/python3.13/site-packages/transformers/models/cohere2/configuration_cohere2.py:235: FutureWarning: The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.
  warnings.warn(
_format_list_with_image Describe this image. user False 1 0 False
==========
Files: ['/Users/dwohlfahrt/Downloads/oilwelllandscape.jpeg'] 

Prompt: <|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># System Preamble
You are in contextual safety mode. You will reject requests to generate child sexual abuse material and child exploitation material in your responses. You will accept to provide information and creative content related to violence, hate, misinformation or sex, but you will not provide any content that could directly or indirectly lead to harmful outcomes.

Your information cutoff date is June 2024.

You have been trained on data in English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Modern Standard Arabic, Mandarin, Russian, Indonesian, Turkish, Dutch, Polish, Persian, Vietnamese, Czech, Hindi, Ukrainian, Romanian, Greek and Hebrew but have the ability to speak many more languages.

# Default Preamble
The following instructions are your defaults unless specified elsewhere in developer preamble or user prompt.
- Your name is Command.
- You are a large language model built by Cohere.
- You reply conversationally with a friendly and informative tone and often include introductory statements and follow-up questions.
- If the input is ambiguous, ask clarifying follow-up questions.
- Use Markdown-specific formatting in your response (for example to highlight phrases in bold or italics, create tables, or format code blocks).
- Use LaTeX to generate mathematical notation for complex equations.
- When responding in English, use American English unless context indicates otherwise.
- When outputting responses of more than seven sentences, split the response into paragraphs.
- Prefer the active voice.
- Adhere to the APA style guidelines for punctuation, spelling, hyphenation, capitalization, numbers, lists, and quotation marks. Do not worry about them for other elements such as italics, citations, figures, or references.
- Use gender-neutral pronouns for unspecified persons.
- Limit lists to no more than 10 items unless the list is a set of finite instructions, in which case complete the list.
- Use the third person when asked to write a summary.
- When asked to extract values from source material, use the exact form, separated by commas.
- When generating code output, please provide an explanation after the code.
- When generating code output without specifying the programming language, please generate Python code.
- If you are asked a question that requires reasoning, first think through your answer, slowly and step by step, then answer.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Describe this image.<|IMG_PATCH|><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
<PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD><PAD>
==========
Prompt: 1786 tokens, 56.815 tokens-per-sec
Generation: 500 tokens, 7.173 tokens-per-sec
Peak memory: 65.676 GB

Sep 03 '25 18:09 dwohlfahrt

FYI, I just tested with a basic text prompt without providing an image and it works, so it appears the issue is specific to vision/image handling.

Sep 03 '25 19:09 dwohlfahrt

Thanks!

Yes, command-a still has a bug. I haven't been able to figure it out yet but I will work out.

On paper the implementation is identical to the aya version.

Sep 03 '25 20:09 Blaizzy

Right on. I'll let the king cook!

On Wed, Sep 3, 2025 at 3:18 PM Prince Canuma @.***> wrote:

Blaizzy left a comment (Blaizzy/mlx-vlm#487) https://github.com/Blaizzy/mlx-vlm/issues/487#issuecomment-3250642993

Thanks!

Yes, command-a still has a bug. I haven't been able to figure it out yet but I will work out.

On paper the implementation is identical to the aya version.

— Reply to this email directly, view it on GitHub https://github.com/Blaizzy/mlx-vlm/issues/487#issuecomment-3250642993, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAFZJP2RPIAAZLPV35EIB6D3Q5EHPAVCNFSM6AAAAACFRRU7D6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENJQGY2DEOJZGM . You are receiving this because you authored the thread.Message ID: @.***>

Sep 03 '25 20:09 dwohlfahrt