github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-24 06:31:50 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #13198] deepseek-ocr via UI only answers if you query it twice #86414

New Issue
Closed
opened 2026-05-10 03:17:19 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-05-10 03:17:19 -05:00
Owner
Copy Link

Originally created by @gparent on GitHub (Nov 21, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/13198

What is the issue?

NOTE: I am not super familiar with Ollama technicals so I'm sorry if there's not enough information in my initial report. I will add whatever is necessary as I learn.

When using the deepseek-ocr model, I have to ask the model a followup question to actually see the reply that was produced in my new chat. This only seems to affect the first prompt to the model.

The behavior that happens is that the model starts "thinking" (triple dots appear), then they suddenly disappear but no answer is shown.

Asking a followup prompt like "What was the answer" then causes the model to give the answer.

Example:

Image

Relevant log output

OS

Windows

GPU

Nvidia

CPU

AMD

Ollama version

0.13.0

Originally created by @gparent on GitHub (Nov 21, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/13198 ### What is the issue? NOTE: I am not super familiar with Ollama technicals so I'm sorry if there's not enough information in my initial report. I will add whatever is necessary as I learn. When using the deepseek-ocr model, I have to ask the model a followup question to actually see the reply that was produced in my new chat. This only seems to affect the first prompt to the model. The behavior that happens is that the model starts "thinking" (triple dots appear), then they suddenly disappear but no answer is shown. Asking a followup prompt like "What was the answer" then causes the model to give the answer. Example: <img width="737" height="260" alt="Image" src="https://github.com/user-attachments/assets/ee7dd4b1-dc5f-4b5b-9c93-77e1dbe2fbe7" /> ### Relevant log output ```shell ``` ### OS Windows ### GPU Nvidia ### CPU AMD ### Ollama version 0.13.0
GiteaMirror added the bug label 2026-05-10 03:17:19 -05:00
GiteaMirror commented 2026-05-10 03:17:22 -05:00
Author
Owner
Copy Link

@gparent commented on GitHub (Nov 21, 2025):

Apologies, I will try to get the log output tonight

<!-- gh-comment-id:3564263066 --> @gparent commented on GitHub (Nov 21, 2025): Apologies, I will try to get the log output tonight
GiteaMirror commented 2026-05-10 03:17:23 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Nov 21, 2025):

Deepseek-OCR does not perform well as a chatbot. It is very sensitive to variations in prompts, and it is recommended to use specific prompts for processing images.

<!-- gh-comment-id:3564277027 --> @rick-github commented on GitHub (Nov 21, 2025): Deepseek-OCR does not perform well as a chatbot. It is very sensitive to variations in prompts, and it is recommended to use [specific prompts](https://github.com/deepseek-ai/DeepSeek-OCR/?tab=readme-ov-file#prompts-examples) for processing images.
GiteaMirror commented 2026-05-10 03:17:23 -05:00
Author
Owner
Copy Link

@gparent commented on GitHub (Nov 21, 2025):

Deepseek-OCR does not perform well as a chatbot. It is very sensitive to variations in prompts, and it is recommended to use specific prompts for processing images.

Thank you! The replies did feel like I was misusing the model, however is it normal to not even get a reply in the first place?

It might not be a bug but it's not a great experience for beginners like me.

Thank you for taking time to explain this, it is appreciated.

<!-- gh-comment-id:3564283018 --> @gparent commented on GitHub (Nov 21, 2025): > Deepseek-OCR does not perform well as a chatbot. It is very sensitive to variations in prompts, and it is recommended to use [specific prompts](https://github.com/deepseek-ai/DeepSeek-OCR/?tab=readme-ov-file#prompts-examples) for processing images. Thank you! The replies did feel like I was misusing the model, however is it normal to not even get a reply in the first place? It might not be a bug but it's not a great experience for beginners like me. Thank you for taking time to explain this, it is appreciated.
GiteaMirror commented 2026-05-10 03:17:24 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Nov 21, 2025):

Notice from your screenshot that DOCR is completing the sentence started with "What was the answer". In this instance, the model is acting as a completion model rather than a chat model. This particular model is trained with a subset of the capabilities of a general purpose model - it responds only to a certain type of input. If you want to use a more chat-oriented model that understands images, I would recommend one of the qwen3-vl models.

<!-- gh-comment-id:3564304693 --> @rick-github commented on GitHub (Nov 21, 2025): Notice from your screenshot that DOCR is completing the sentence started with "What was the answer". In this instance, the model is acting as a completion model rather than a chat model. This particular model is trained with a subset of the capabilities of a general purpose model - it responds only to a certain type of input. If you want to use a more chat-oriented model that understands images, I would recommend one of the [qwen3-vl](https://ollama.com/library/qwen3-vl) models.
GiteaMirror commented 2026-05-10 03:17:24 -05:00
Author
Owner
Copy Link

@gparent commented on GitHub (Nov 21, 2025):

Notice from your screenshot that DOCR is completing the sentence started with "What was the answer". In this instance, the model is acting as a completion model rather than a chat model. This particular model is trained with a subset of the capabilities of a general purpose model - it responds only to a certain type of input. If you want to use a more chat-oriented model that understands images, I would recommend one of the qwen3-vl models.

Thanks! Yeah I also noticed that if I don't add a question mark, it starts the answers with one. I planned to use deepseek-ocr programmatically in the first place to handle some bills of mine. I think improving the way I use it (using correct prompts and not chatting with it) is going to be the way to go, but I'll keep qwen3-vl in mind for interactive usage.

Again I cannot overstate how much I appreciate you taking time to explain this to me. I don't think this is really a bug anymore, other than the possible UI experience quirk for people like me who do not know better.

<!-- gh-comment-id:3564311736 --> @gparent commented on GitHub (Nov 21, 2025): > Notice from your screenshot that DOCR is completing the sentence started with "What was the answer". In this instance, the model is acting as a completion model rather than a chat model. This particular model is trained with a subset of the capabilities of a general purpose model - it responds only to a certain type of input. If you want to use a more chat-oriented model that understands images, I would recommend one of the [qwen3-vl](https://ollama.com/library/qwen3-vl) models. Thanks! Yeah I also noticed that if I don't add a question mark, it starts the answers with one. I planned to use deepseek-ocr programmatically in the first place to handle some bills of mine. I think improving the way I use it (using correct prompts and not chatting with it) is going to be the way to go, but I'll keep qwen3-vl in mind for interactive usage. Again I cannot overstate how much I appreciate you taking time to explain this to me. I don't think this is really a bug anymore, other than the possible UI experience quirk for people like me who do not know better.
GiteaMirror commented 2026-05-10 03:17:25 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Nov 21, 2025):

An example of programmatic use of this model:

#!/usr/bin/env python3

import ollama
import sys

model = 'deepseek-ocr:3b-bf16'
image = sys.argv[1] if len(sys.argv) > 1 else "puppy.jpg"

response = ollama.chat(
    model=model,
    messages=[{
        'role': 'user',
        'content': '\nDescribe this image.',
        'images': [image]
    }],
    stream=True
)

for r in response:
  print(r['message']['content'], end='')
print("")
<!-- gh-comment-id:3564316439 --> @rick-github commented on GitHub (Nov 21, 2025): An example of programmatic use of this model: ```python #!/usr/bin/env python3 import ollama import sys model = 'deepseek-ocr:3b-bf16' image = sys.argv[1] if len(sys.argv) > 1 else "puppy.jpg" response = ollama.chat( model=model, messages=[{ 'role': 'user', 'content': '\nDescribe this image.', 'images': [image] }], stream=True ) for r in response: print(r['message']['content'], end='') print("") ```
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#86414
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?