diff --git a/assistant/assistant.py b/assistant/assistant.py
index 5be7ebe..1618b25 100644
--- a/assistant/assistant.py
+++ b/assistant/assistant.py
@@ -45,6 +45,7 @@ class Assistant(commands.Cog):
             enabled=True,
             chat_channel=None,
             model="gemma4",
+            vision_model="gemma4",
             temperature=0.7,
             system_prompt="You are a helpful assistant.",
             question_mode=False,
@@ -96,7 +97,29 @@ class Assistant(commands.Cog):
             for mention in (f"<@{self.bot.user.id}>", f"<@!{self.bot.user.id}>"):
                 content = content.replace(mention, "").strip()
 
-        if not content:
+        # Collect image URLs from this message and any replied-to message
+        image_urls = []
+        for attachment in message.attachments:
+            if attachment.content_type and attachment.content_type.startswith("image/"):
+                image_urls.append(attachment.url)
+
+        # Check replied-to message for images
+        if message.reference and message.reference.message_id:
+            try:
+                ref_msg = await message.channel.fetch_message(message.reference.message_id)
+                for attachment in ref_msg.attachments:
+                    if attachment.content_type and attachment.content_type.startswith("image/"):
+                        image_urls.append(attachment.url)
+                # Also check embeds for image URLs
+                for embed in ref_msg.embeds:
+                    if embed.image and embed.image.url:
+                        image_urls.append(embed.image.url)
+                    if embed.thumbnail and embed.thumbnail.url:
+                        image_urls.append(embed.thumbnail.url)
+            except (discord.NotFound, discord.HTTPException):
+                pass
+
+        if not content and not image_urls:
             return
 
         if question_mode and not content.rstrip().endswith("?"):
@@ -105,7 +128,7 @@ class Assistant(commands.Cog):
             return
 
         async with message.channel.typing():
-            reply = await self._chat(guild, message.channel.id, content)
+            reply = await self._chat(guild, message.channel.id, content, image_urls=image_urls)
 
         if reply is None:
             await message.reply("No response from the assistant.")
@@ -126,7 +149,7 @@ class Assistant(commands.Cog):
     # API
     # ---------------------------------------------------------------------------
 
-    async def _chat(self, guild: discord.Guild, channel_id: int, user_msg: str) -> str | None:
+    async def _chat(self, guild: discord.Guild, channel_id: int, user_msg: str, image_urls: list[str] | None = None) -> str | None:
         api_key = await self.config.api_key()
         if not api_key or not self.session:
             return None
@@ -137,11 +160,27 @@ class Assistant(commands.Cog):
         max_tokens = await self.config.guild(guild).max_tokens()
         reasoning_effort = await self.config.guild(guild).reasoning_effort()
 
+        # Auto-switch to vision model when images are present
+        if image_urls:
+            vision_model = await self.config.guild(guild).vision_model()
+            if vision_model:
+                model = vision_model
+
         key = (guild.id, channel_id)
         if key not in self.history:
             self.history[key] = []
 
-        self.history[key].append({"role": "user", "content": user_msg})
+        # Build user message content (text-only or multimodal)
+        if image_urls:
+            # OpenAI vision format: content is an array of parts
+            user_content = []
+            if user_msg:
+                user_content.append({"type": "text", "text": user_msg})
+            for url in image_urls:
+                user_content.append({"type": "image_url", "image_url": {"url": url}})
+            self.history[key].append({"role": "user", "content": user_content})
+        else:
+            self.history[key].append({"role": "user", "content": user_msg})
 
         messages = [{"role": "system", "content": system_prompt}] + self.history[key][-20:]
 
@@ -231,6 +270,20 @@ class Assistant(commands.Cog):
         await self.config.guild(ctx.guild).model.set(model)
         await ctx.send(f"Model set to **{model}**.")
 
+    @assistant.command(name="visionmodel")
+    async def assistant_visionmodel(self, ctx: commands.Context, model: str = None):
+        """Set the model used when images are detected. Omit to show current.
+
+        Vision-capable models: gemma4, green-l, green-r, gpt-oss-120b,
+        gemma-3-27b-it, mistral-small-3.2-24b-instruct-2506, kimi-k2.6, kimi-k3
+        """
+        if model is None:
+            current = await self.config.guild(ctx.guild).vision_model()
+            await ctx.send(f"Vision model: **{current}**")
+        else:
+            await self.config.guild(ctx.guild).vision_model.set(model)
+            await ctx.send(f"Vision model set to **{model}**.")
+
     @assistant.command(name="models")
     async def assistant_models(self, ctx: commands.Context):
         """List available chat models sorted by cost (cheapest first)."""
@@ -414,6 +467,7 @@ class Assistant(commands.Cog):
         enabled = await cfg.enabled()
         chat_channel = await cfg.chat_channel()
         model = await cfg.model()
+        vision_model = await cfg.vision_model()
         temp = await cfg.temperature()
         sys_prompt = await cfg.system_prompt()
         q_mode = await cfg.question_mode()
@@ -428,6 +482,7 @@ class Assistant(commands.Cog):
         embed.add_field(name="Chat Channel", value=channel_str, inline=True)
         embed.add_field(name="API Key", value="Set" if has_key else "Not set", inline=True)
         embed.add_field(name="Model", value=model, inline=True)
+        embed.add_field(name="Vision Model", value=vision_model, inline=True)
         embed.add_field(name="Temperature", value=str(temp), inline=True)
         embed.add_field(name="Question Mode", value=str(q_mode), inline=True)
         embed.add_field(name="Max Tokens", value=str(max_tokens) if max_tokens else "Unlimited", inline=True)