Unable to read images for some models

Description

When I paste an image into opencode, it keeps telling me that it cannot read images, after it has been attached. On version 1.0.134 it works, but not on 1.0.137 onwards.

This is the stack

Backend: LiteLLM + Vertex AI

OpenCode version

1.0.143

Steps to reproduce

Screenshot an image
Paste into opencode

Screenshot and/or share link

No response

Operating System

macOS 26.1 Tahoe

Terminal

VSCode

This issue might be a duplicate of existing issues. Please check:

opus not able to read images with read tool #5056: opus not able to read images with read tool
"I cannot analyze or describe image files. ", "The read tool cannot process binary files like images." when trying to use image output file from previous turn #2825: "I cannot analyze or describe image files" error when trying to use image output files
Unsure if image analysis is supported in OpenCode (Windows 11, WezTerm & PowerShell) #2937: Image analysis support issues on Windows

Feel free to ignore if none of these address your specific case.

I looked further into the issue and found that it works if I set the input modality to include image. So I think this might be an issue with the detection of models + modalities.

The particular case I faced: claude-sonnet-4-5-20250929 was not being detected as a model that could support images.

For those interested, the property to set was:

{
  "<model>": {
    //...
    "modalities": { "input": ["image", "text"], "output": ["text"] }
  }
{

was it custom or is it one that we show in models.dev

Is there a setting on opencode to set "default_image_model" to handle image inputs when the current model is not able to process it ?

there isnt no

I get this error when using us.anthropic.claude-haiku-4-5-20251001-v1:0 and even us.anthropic.claude-sonnet-4-5-20250929-v1:0 but using global.anthropic.claude-opus-4-5-20251101-v1:0 works fine. Weirdly, when I switch back from Opus to Sonnet, it is able to read the image again.

@MrBazzieB

@MrBazzieB what error are you seeing>

@rekram1-node

@rekram1-node Is this something that the opencode team would consider ? In similar veins to the "small_model" having a "vision_model" could be interesting too

I think u can accomplish this via plugins, make a hook for on tool output, check if it has attachments, drop them from the tool call and instead call an llm to explain the image and say what it is or something

but yeah maybewe need something better for that behavior

@rekram1-node

@MrBazzieB what error are you seeing>

@rekram1-node I am not seeing any error, other than this message:

I'm unable to read image files directly as I don't have image input capabilities. However, I can help you in other ways:
1. Use an OCR tool - You could use an external tool to extract text from the image and paste it here
2. Describe the error - Tell me what error you're seeing and I can help troubleshoot
3. Check the file - Since the image is at /path/to/image.png, it seems to be related to a report error in your application
Could you describe what error is shown in the image, or paste any error messages you're seeing? That way I can help you fix the issue.
▣  Build · us.anthropic.claude-sonnet-4-5-20250929-v1:0 · 7.3s

When I switch to OPUS, it is able to read the image, and when I switch back to SONNET it is able to read the image. I then went ahead and got OPENCODE to debug the issue and search the logs, this is what it told me:

Based on the logs, I found the issue. The error message about the image appears to be generated client-side by >OpenCode, not from the actual logs. Here's what I found:
What happened with the image:
The log shows that when you tried to attach the image /path/to/image.png, OpenCode detected that:

The model being used (us.anthropic.claude-sonnet-4-5-20250929-v1:0 via Amazon Bedrock) does not support image >input

OpenCode intercepted this before sending the request to the API

It displayed the error message: ERROR: Cannot read "/path/to/image.png" (this model does not support image input)
Why this happened:
The error log shows AI_APICallError entries around the same timestamp (2025-12-23T15:34:16 and 2025-12->23T20:03:41), indicating the model was trying to process your request but couldn't handle the image attachment.
Solution:
To share images with me, you have a few options:

Describe the error - Tell me what you see in the image

Use a vision-capable model - Switch to a model that supports images (like Claude Sonnet 4 via Anthropic direct, or GPT->4 Vision)

Extract text - Use OCR to extract any text from the image and paste it here
What error are you seeing in the /path/to/image.png image? I'm happy to help troubleshoot once you describe

I looked further into the issue and found that it works if I set the input modality to include image. So I think this might be an issue with the detection of models + modalities.

The particular case I faced: claude-sonnet-4-5-20250929 was not being detected as a model that could support images.

For those interested, the property to set was:

{
"": {
//...
"modalities": { "input": ["image", "text"], "output": ["text"] }
}
{

Thanks a bunch this got Qwen 3.5 on LM studio to start reading images as well.

@kenthinson

I looked further into the issue and found that it works if I set the input modality to include image. So I think this might be an issue with the detection of models + modalities.我进一步调查了这个问题，发现如果我把输入模态设置为包含图片，它就能正常工作。所以我认为这可能是模型+模态检测的问题。
The particular case I faced: claude-sonnet-4-5-20250929 was not being detected as a model that could support images.我遇到的具体案例是：claude-sonnet-4-5-20250929 没有被检测到支持图像的模型。
For those interested, the property to set was:有兴趣的人，要设定的属性是：
{
"": {
//...
"modalities": { "input": ["image", "text"], "output": ["text"] }“模态”： { “输入”： [“图像”， “文本”]， “输出”： [“文本”] }
}
{

Thanks a bunch this got Qwen 3.5 on LM studio to start reading images as well.非常感谢，这让 LM Studio 上的 Qwen 3.5 也能开始读取图像了。

@kenthinson How can I enable a multimodal model to understand images in opencode? In my tests, the LLM received base64 encoding, or did it fail to understand images correctly

Im using gemma 4 with custom provider and modalities option didn't work. But I was missing typing "npm": "@ai-sdk/openai-compatiple" in provider, and it worked after configuring it.

por que si este modelo dice qu epuede leer images y le mando una imagen y entonces me dice que no lo puede hacer? yo pague pensando en que sus modelos eran capaces de leer imágenes cuando no es asi. es una estafa?

por que si este modelo dice qu epuede leer images y le mando una imagen y entonces me dice que no lo puede hacer? yo pague pensando en que sus modelos eran capaces de leer imágenes cuando no es asi. es una estafa?

Im using gemma 4 with custom provider and modalities option didn't work. But I was missing typing "npm": "@ai-sdk/openai-compatiple" in provider, and it worked after configuring it.

It seems like the image understanding of capable models is still not used for all providers in OpenCode. I've tested this with various Ollama models capable of image usage. Other software tools (e.g. Open WebUI) or a simple Python script are also able to use the models capabilities. I do not understand why it is not available by default in OpenCode though...

The setup that worked for me was to use "modalities": { "input": ["image", "text"], "output": ["text"] } in the opencode.json alongside with the "npm": "@ai-sdk/openai-compatible".

Here is my config json example where ONLY gemma4:12b-it-q4_K_M works as it has the modalities addition:

{
	"$schema": "https://opencode.ai/config.json",
	"provider": {
		"ollama": {
			"models": {
				"qwen3.6:27b-q4_K_M": {
					"_launch": true,
					"name": "qwen3.6:27b-q4_K_M"
				},
				"gemma4:26b-a4b-it-q4_K_M": {
					"name": "gemma4:26b-a4b-it-q4_K_M"
				},
				"gemma4:12b-it-q4_K_M": {
					"name": "gemma4:12b-it-q4_K_M",
                                        "modalities": { "input": ["image", "text"], "output": ["text"] }
				},
				"gemma4:e4b-it-q4_K_M": {
					"name": "gemma4:e4b-it-q4_K_M"
				},
				"qwen3-vl:8b-instruct-q4_K_M": {
					"name": "qwen3-vl:8b-instruct-q4_K_M"
				},
				"ministral-3:14b-instruct-2512-q4_K_M": {
					"name": "ministral-3:14b-instruct-2512-q4_K_M"
				}
				
			},
			"name": "Ollama",
			"npm": "@ai-sdk/openai-compatible",
			"options": {
				"baseURL": "http://127.0.0.1:11434/v1"
			}
		}
	}
}

If you would want to use images with the other models, you would need to add "modalities": { "input": ["image", "text"], "output": ["text"] } , too.

For anyone using an Ollama instance, here's a simple Python script for you to test if the model supports image input:
However, you could also have a look at the model description on https://ollama.com/library/<YOUR-MODEL>/tags

from tkinter import Tk, filedialog

import ollama
from rich.console import Console
from rich.markdown import Markdown

def image_vision():

    root = Tk()
    root.withdraw()

    image_path = filedialog.askopenfilename(defaultextension=".jpg", filetypes=[("Image files", "*.jpg"), ("Image files", "*.png"), ("Image files", "*.jpeg")])

    response = ollama.chat(
        model='YOUR-MODEL-NAME', # e.g. gemma4:12b-it-q4_K_M
        messages=[{
            'role': 'user',
            'content': 'What is in this image? If it is text extract the text and write it as Markdown. If there is no text describe the image. Answer in Spanish.',
            'images': [image_path]
        }]
    )

    console = Console()
    console.print(response)
    console.print(Markdown("## Response"))
    console.print(Markdown(response['message']['content']))


def main():
    image_vision()

if __name__ == "__main__":
    main()

Which results in:

Response                                                                                                                                                                                                                                                                                                                                                            
Esta es una vista aérea o satelital de una zona urbana densamente construida, posiblemente una ciudad europea debido al estilo arquitectónico y los techos de teja roja.                                                                                                                                                                                            

En la parte superior central, se observa un gran parque con amplias áreas verdes, árboles y lo que parece ser un camino circular. A la derecha del parque, hay un campo deportivo rectangular con una pista de atletismo de color rojo. El resto de la imagen muestra una red de edificios compactos organizados a lo largo de calles y avenidas. Se pueden         
distinguir diversas estructuras residenciales y comerciales, pequeños patios internos y zonas arboladas intercaladas entre las edificaciones.

Workaround : plugin opencode qui donne la vision aux modèles sans capacité image (testé avec deepseek-v4-flash sous OpenChamber 1.18.15, 07/08/2026).

Le problème exact décrit ici : quand on colle une image, opencode la stocke en base64 dans la DB (aucun fichier sur disque) et un modèle sans vision ne peut pas la traiter. Mon plugin vision-attachments intercepte le message utilisateur (hook chat.message), écrit l'image sur disque (/tmp/opencode-vision/), retire la part image et injecte une note texte avec le chemin → l'agent appelle automatiquement l'outil analyze_image d'un MCP vision externe (ex: opencode-vision avec Gemini 2.5 Flash free tier) qui renvoie une description texte lisible par le modèle principal.

→ https://github.com/zagg/opencode-vision-attachments

Pièges documentés dans le README (muter output.parts par splice au lieu de réassigner, hériter id/messageID/sessionID de la part remplacée, ne pas throw dans le hook). En espérant qu'un support natif arrive un jour 👍

@rekram1-node

@rekram1-node J'ai implémenté exactement l'approche que tu avais esquissée ici (hook, drop des attachments, appel d'un LLM vision externe pour décrire l'image). Plugin opencode testé avec deepseek-v4-flash sous OpenChamber 1.18.15 : https://github.com/zagg/opencode-vision-attachments — les pièges (muter output.parts par splice, hériter les IDs) sont documentés dans le README. 👍

Closing as stale. This was reported as a regression in v1.0.137 with LiteLLM + Vertex AI, and model and attachment handling has been rewritten since then. Later comments here describe a different case: models that don't support images.

If images still fail on the latest version with a model that supports vision, please open a new issue with your provider config, model, and version.

Navigation Menu

Repository navigation

Unable to read images for some models #5359

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Unable to read images for some models #5359

Description

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Dec 11, 2025

conradkoh commented on Dec 11, 2025

rekram1-node commented on Dec 11, 2025

mitermayer commented on Dec 17, 2025

rekram1-node commented on Dec 17, 2025

MrBazzieB commented on Dec 17, 2025

rekram1-node commented on Dec 18, 2025

mitermayer commented on Dec 21, 2025

rekram1-node commented on Dec 21, 2025

rekram1-node commented on Dec 21, 2025

MrBazzieB commented on Dec 24, 2025

kenthinson commented on Mar 8, 2026

narcissus1024 commented on Mar 10, 2026

kim-youngseop-developer commented on Apr 3, 2026

kelvin2403901 commented on Jun 6, 2026

CodeMazeSolver commented on Jul 6, 2026

zagg commented on Aug 7, 2026

zagg commented on Aug 7, 2026

neriousy commented on Oct 1, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions