diff --git a/app/frontend/src/components/ModelManager.jsx b/app/frontend/src/components/ModelManager.jsx
index 1c0f44c..e62d14e 100644
--- a/app/frontend/src/components/ModelManager.jsx
+++ b/app/frontend/src/components/ModelManager.jsx
@@ -35,6 +35,7 @@ const MODEL_LIBRARY = [
         resolution: "1024x1024",
         notes: "Photorealistic SDXL model tuned for 4-8 step Lightning generation.",
         url: "https://huggingface.co/RunDiffusion/Juggernaut-XL-Lightning/resolve/main/Juggernaut_RunDiffusionPhoto2_Lightning_4Steps.safetensors",
+        recommendedTiers: ["Mid", "High"],
       },
       {
         name: "DreamShaper XL Lightning",
@@ -44,6 +45,7 @@ const MODEL_LIBRARY = [
         resolution: "1024x1024",
         notes: "All-around SDXL model for fantasy art, renders, illustration, and stylized images.",
         url: "https://huggingface.co/Lykon/dreamshaper-xl-lightning/resolve/main/DreamShaperXL_Lightning.safetensors",
+        recommendedTiers: ["Mid", "High"],
       },
     ],
   },
@@ -58,6 +60,7 @@ const MODEL_LIBRARY = [
         resolution: "512x512",
         notes: "General-purpose SD 1.5 model for illustration, anime, and semi-realistic portraits.",
         url: "https://huggingface.co/Lykon/DreamShaper/resolve/main/DreamShaper_8_pruned.safetensors",
+        recommendedTiers: ["Low"],
       },
       {
         name: "CyberRealistic V8",
@@ -67,6 +70,7 @@ const MODEL_LIBRARY = [
         resolution: "512x512",
         notes: "High-fidelity realism and photorealism in the SD 1.5 ecosystem.",
         url: "https://huggingface.co/cyberdelia/CyberRealistic/resolve/main/CyberRealistic_V8_FP16.safetensors",
+        recommendedTiers: ["Low"],
       },
       {
         name: "ReV Animated v1.2.2",
@@ -76,6 +80,7 @@ const MODEL_LIBRARY = [
         resolution: "512x512",
         notes: "Semi-realistic, 2.5D, anime, fantasy, and stylized digital art.",
         url: "https://huggingface.co/danbrown/RevAnimated-v1-2-2/resolve/main/rev-animated-v1-2-2.safetensors",
+        recommendedTiers: ["Low"],
       },
     ],
   },
@@ -93,6 +98,7 @@ const TEXT_MODEL_LIBRARY = [
         resolution: "N/A",
         notes: "Extremely fast, lightweight assistant, perfect for low RAM/VRAM machines.",
         url: "https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct-GGUF/resolve/main/qwen2.5-coder-0.5b-instruct-q4_k_m.gguf",
+        recommendedTiers: ["Low"],
       },
       {
         name: "SmolLM2 1.7B Instruct Q4_K_M",
@@ -102,6 +108,42 @@ const TEXT_MODEL_LIBRARY = [
         resolution: "N/A",
         notes: "Excellent lightweight assistant with strong logic, reasoning, and prompt expansion capabilities.",
         url: "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF/resolve/main/smollm2-1.7b-instruct-q4_k_m.gguf",
+        recommendedTiers: ["Low"],
+      },
+      {
+        name: "Qwen2.5-Coder-7B-Instruct",
+        filename: "qwen2.5-coder-7b-instruct-q4_k_m.gguf",
+        format: "GGUF",
+        size: "4.7 GB",
+        approxSize: "4.7 GB",
+        resolution: "N/A",
+        notes: "Highly intelligent coding and text assistant. Recommended for mid/high tier systems.",
+        url: "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/resolve/main/qwen2.5-coder-7b-instruct-q4_k_m.gguf",
+        recommendedTiers: ["Mid", "High"],
+      },
+      {
+        name: "Llama-3.1-8B-Instruct",
+        filename: "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
+        format: "GGUF",
+        size: "4.9 GB",
+        approxSize: "4.9 GB",
+        resolution: "N/A",
+        notes: "Excellent general-purpose text model. Recommended for high-tier systems.",
+        url: "https://huggingface.co/lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
+        recommendedTiers: ["High"],
+      },
+      {
+        name: "LLaVA 1.5 7B",
+        filename: "ggml-model-q4_k.gguf",
+        format: "GGUF",
+        size: "4.5 GB",
+        approxSize: "4.5 GB",
+        resolution: "N/A",
+        notes: "Multimodal model capable of understanding images. Downloads companion vision projector automatically.",
+        url: "https://huggingface.co/mys/ggml_llava-v1.5-7b/resolve/main/ggml-model-q4_k.gguf",
+        projectorUrl: "https://huggingface.co/mys/ggml_llava-v1.5-7b/resolve/main/mmproj-model-f16.gguf",
+        projectorFilename: "mmproj-model-f16.gguf",
+        recommendedTiers: ["Mid", "High"],
       },
     ],
   },
@@ -143,6 +185,20 @@ const COREML_MODEL_LIBRARY = [
   },
 ];
 
+const getHardwareTier = (specs) => {
+  if (!specs) return "Low";
+  const ram = Number(specs.ram_total_gb) || 0;
+  const vram = Number(specs.gpu_vram_gb) || 0;
+
+  if (vram >= 12 || (vram >= 8 && ram >= 32)) {
+    return "High";
+  } else if (vram >= 6 || ram >= 16) {
+    return "Mid";
+  } else {
+    return "Low";
+  }
+};
+
 function ModelManager({ 
   activeModel, 
   setActiveModel, 
@@ -288,32 +344,6 @@ function ModelManager({
     }
   };
 
-  const startProgressPolling = (modelId) => {
-    setDownloadingModelId(modelId);
-    
-    const interval = setInterval(async () => {
-      try {
-        const status = await getDownloadProgress();
-        if (status.active) {
-          setDownloadProgress(status.progress === -1 ? 0 : status.progress);
-          setDownloadSpeed(`${status.speed} • ETA ${status.eta}s`);
-        } else {
-          clearInterval(interval);
-          setDownloadingModelId(null);
-          if (status.error) {
-            if (!String(status.error).toLowerCase().includes("cancelled")) {
-              showAlert({ title: "Download Failed", message: status.error, danger: true });
-            }
-          } else {
-            fetchModels();
-          }
-        }
-      } catch (e) {
-        console.error("Polling error:", e);
-      }
-    }, 1000);
-  };
-
   const downloadByUrl = async (url, expectedFilename) => {
     if (downloadingModelId) return;
 
@@ -352,6 +382,40 @@ function ModelManager({
     }
   };
 
+  const startProgressPolling = (modelId) => {
+    setDownloadingModelId(modelId);
+    
+    const interval = setInterval(async () => {
+      try {
+        const status = await getDownloadProgress();
+        if (status.active) {
+          setDownloadProgress(status.progress === -1 ? 0 : status.progress);
+          setDownloadSpeed(`${status.speed} • ETA ${status.eta}s`);
+        } else {
+          clearInterval(interval);
+          setDownloadingModelId(null);
+          if (status.error) {
+            if (!String(status.error).toLowerCase().includes("cancelled")) {
+              showAlert({ title: "Download Failed", message: status.error, danger: true });
+            }
+          } else {
+            fetchModels();
+            if (modelId === "ggml-model-q4_k.gguf") {
+              setTimeout(() => {
+                downloadByUrl(
+                  "https://huggingface.co/mys/ggml_llava-v1.5-7b/resolve/main/mmproj-model-f16.gguf",
+                  "mmproj-model-f16.gguf"
+                );
+              }, 500);
+            }
+          }
+        }
+      } catch (e) {
+        console.error("Polling error:", e);
+      }
+    }, 1000);
+  };
+
   const handleUrlDownloadSubmit = async (e) => {
     e.preventDefault();
     if (!downloadUrl.trim()) return;
@@ -809,6 +873,8 @@ function ModelManager({
         </button>
       </div>
 
+
+
       {/* Active Model Status Tonal Box */}
       {activeModelType === "image" && activeModel && (
         <div className="m3-card" style={{ borderLeft: "4px solid var(--md-sys-color-primary)", background: "var(--md-sys-color-primary-container)", color: "var(--md-sys-color-on-primary-container)" }}>
@@ -1006,10 +1072,27 @@ function ModelManager({
               {section.items.map((model) => {
                 const installed = modelNames.includes(model.filename);
                 const downloading = downloadingModelId === model.filename || downloadingModelId === `${model.filename}.zip`;
+                const systemTier = getHardwareTier(specs);
+                const isRecommended = model.recommendedTiers && model.recommendedTiers.includes(systemTier);
                 return (
                   <div key={model.filename} style={{ display: "flex", flexDirection: "column", gap: "10px", padding: "14px", background: "var(--md-sys-color-surface-variant)", border: "1px solid var(--md-sys-color-outline-variant)", borderRadius: "var(--md-shape-corner-medium)" }}>
                     <div style={{ display: "flex", flexDirection: "column", gap: "4px" }}>
-                      <span style={{ fontWeight: 700 }}>{model.name}</span>
+                      <span style={{ fontWeight: 700, display: "flex", alignItems: "center", gap: "6px", flexWrap: "wrap" }}>
+                        {model.name}
+                        {isRecommended && (
+                          <span style={{ 
+                            fontSize: "0.68rem", 
+                            color: "#2e7d32", 
+                            background: "#e8f5e9", 
+                            padding: "2px 6px", 
+                            borderRadius: "4px", 
+                            fontWeight: "bold",
+                            border: "1px solid #c8e6c9" 
+                          }}>
+                            [Recommended Fit]
+                          </span>
+                        )}
+                      </span>
                       <span style={{ fontSize: "0.75rem", color: "var(--md-sys-color-outline)" }}>
                         {model.format} • approx. {model.approxSize} {model.resolution !== "N/A" && `• ${model.resolution}`}
                       </span>
diff --git a/app/frontend/src/components/TextChat.jsx b/app/frontend/src/components/TextChat.jsx
index c03edc3..a697dfa 100644
--- a/app/frontend/src/components/TextChat.jsx
+++ b/app/frontend/src/components/TextChat.jsx
@@ -1,5 +1,5 @@
 import React, { useCallback, useEffect, useRef, useState } from "react";
-import { Bot, LoaderCircle, Send, Trash2, Square, History, Plus } from "lucide-react";
+import { Bot, LoaderCircle, Send, Trash2, Square, History, Plus, Paperclip, X } from "lucide-react";
 import {
   chatWithLlm,
   getDownloadProgress,
@@ -30,6 +30,64 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
   const completedDownloadRef = useRef("");
   const loadingModelRef = useRef(null);
 
+  const [attachments, setAttachments] = useState([]);
+  const fileInputRef = useRef(null);
+
+  const isImage = (file) => {
+    return /\.(jpe?g|png|webp)$/i.test(file.name) || file.type.startsWith("image/");
+  };
+
+  const isTextFile = (file) => {
+    return /\.(txt|md|csv|js|jsx|ts|tsx|py|json|css|html|java|cpp|c|h|rs|go|sh|bat|xml|yaml|yml)$/i.test(file.name) || file.type.startsWith("text/");
+  };
+
+  const handleFileChange = (e) => {
+    const files = Array.from(e.target.files);
+    if (!files || files.length === 0) return;
+
+    files.forEach((file) => {
+      if (isImage(file)) {
+        const reader = new FileReader();
+        reader.onload = (event) => {
+          setAttachments((prev) => [
+            ...prev,
+            {
+              id: Math.random().toString(36).substring(7),
+              file,
+              type: "image",
+              name: file.name,
+              dataUrl: event.target.result,
+            },
+          ]);
+        };
+        reader.readAsDataURL(file);
+      } else if (isTextFile(file)) {
+        const reader = new FileReader();
+        reader.onload = (event) => {
+          setAttachments((prev) => [
+            ...prev,
+            {
+              id: Math.random().toString(36).substring(7),
+              file,
+              type: "document",
+              name: file.name,
+              content: event.target.result,
+            },
+          ]);
+        };
+        reader.readAsText(file);
+      } else {
+        showAlert({
+          title: "Unsupported File",
+          message: `File "${file.name}" is not supported. Please select an image (JPEG, PNG, WEBP) or a text/code file.`,
+          danger: true,
+        });
+      }
+    });
+
+    e.target.value = "";
+  };
+
   useEffect(() => {
     loadingModelRef.current = loadingModel;
   }, [loadingModel]);
@@ -179,7 +237,12 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
       setSelectedModel(conv.model);
     }
     // Try to compute approximate token usage from messages
-    const total = conv.messages.reduce((sum, m) => sum + m.content.split(/\s+/).length, 0);
+    const total = conv.messages.reduce((sum, m) => {
+      const text = Array.isArray(m.content)
+        ? m.content.map(c => c.text || "").join(" ")
+        : (m.content || "");
+      return sum + text.split(/\s+/).length;
+    }, 0);
     setTokenUsage({
       prompt_tokens: Math.round(total * 0.7),
       completion_tokens: Math.round(total * 0.3),
@@ -204,7 +267,8 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
 
   const sendMessage = async () => {
     const text = input.trim();
-    if (!text || isBusy || !status.ready) return;
+    const hasAttachments = attachments.length > 0;
+    if ((!text && !hasAttachments) || isBusy || !status.ready) return;
 
     let convId = activeConversationId;
     let isNew = false;
@@ -214,12 +278,40 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
       isNew = true;
     }
 
-    const nextMessages = [...messages, { role: "user", content: text }];
+    // Delimited context blocks for document attachments
+    let documentContext = "";
+    attachments.filter(att => att.type === "document").forEach((att) => {
+      documentContext += `\n[Attached File: ${att.name}]\n${att.content}\n`;
+    });
+
+    const combinedText = documentContext ? `${text}\n\n${documentContext}`.trim() : text;
+    const imageAttachments = attachments.filter(att => att.type === "image");
+
+    let userMessageContent;
+    if (imageAttachments.length > 0) {
+      userMessageContent = [
+        {
+          type: "text",
+          text: combinedText
+        },
+        ...imageAttachments.map((img) => ({
+          type: "image_url",
+          image_url: {
+            url: img.dataUrl
+          }
+        }))
+      ];
+    } else {
+      userMessageContent = combinedText;
+    }
+
+    const nextMessages = [...messages, { role: "user", content: userMessageContent }];
     setMessages(nextMessages);
     setInput("");
     setIsBusy(true);
 
-    const firstTitle = isNew ? (text.slice(0, 26) + (text.length > 26 ? "..." : "")) : null;
+    const displayTitleText = text || (imageAttachments.length > 0 ? "Sent Image" : "Sent File");
+    const firstTitle = isNew ? (displayTitleText.slice(0, 26) + (displayTitleText.length > 26 ? "..." : "")) : null;
     saveConversationState(convId, nextMessages, selectedModel, firstTitle);
 
     try {
@@ -238,6 +330,9 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
         setTokenUsage(response.usage);
       }
       saveConversationState(convId, finalMessages, selectedModel);
+      
+      // Clean up attached files on success
+      setAttachments([]);
     } catch (err) {
       const finalMessages = [...nextMessages, { role: "assistant", content: `Error: ${err.message}`, error: true }];
       setMessages(finalMessages);
@@ -510,7 +605,30 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
               {messages.map((message, index) => (
                 <div key={`${message.role}-${index}`} className={`chat-message ${message.role} ${message.error ? "error" : ""}`}>
                   <strong>{message.role === "user" ? "You" : "Local AI"}</strong>
-                  <div>{message.content}</div>
+                  <div>
+                    {Array.isArray(message.content) ? (
+                      <div style={{ display: "flex", flexDirection: "column", gap: "8px" }}>
+                        {message.content.map((item, idx) => {
+                          if (item.type === "text") {
+                            return <span key={idx} style={{ whiteSpace: "pre-wrap" }}>{item.text}</span>;
+                          }
+                          if (item.type === "image_url") {
+                            return (
+                              <img 
+                                key={idx} 
+                                src={item.image_url.url} 
+                                alt="Attached image" 
+                                style={{ maxWidth: "240px", maxHeight: "180px", objectFit: "contain", borderRadius: "6px", marginTop: "4px" }} 
+                              />
+                            );
+                          }
+                          return null;
+                        })}
+                      </div>
+                    ) : (
+                      <span style={{ whiteSpace: "pre-wrap" }}>{message.content}</span>
+                    )}
+                  </div>
                 </div>
               ))}
               {isBusy && status.ready && <div className="chat-thinking"><LoaderCircle className="progress-spinner" size={16} /> Generating...</div>}
@@ -519,22 +637,99 @@ function TextChat({ specs, showAlert, showConfirm, textSettings, setTextSettings
           <div ref={bottomRef} />
         </div>
 
-        <div className="chat-composer">
-          <textarea
-            value={input}
-            onChange={(event) => setInput(event.target.value)}
-            onKeyDown={(event) => {
-              if (event.key === "Enter" && !event.shiftKey) {
-                event.preventDefault();
-                sendMessage();
-              }
-            }}
-            placeholder={status.ready ? "Message your local model..." : "Select and load a GGUF model above to begin"}
-            disabled={!status.ready || isBusy}
-          />
-          <button className="m3-btn m3-btn-filled" onClick={sendMessage} disabled={!input.trim() || !status.ready || isBusy}>
-            <Send size={17} /> Send
-          </button>
+        <div className="chat-composer" style={{ display: "flex", flexDirection: "column", gap: "10px" }}>
+          {attachments.length > 0 && (
+            <div style={{ display: "flex", flexWrap: "wrap", gap: "8px", paddingBottom: "8px" }}>
+              {attachments.map((att) => (
+                <div key={att.id} style={{ 
+                  position: "relative", 
+                  display: "flex", 
+                  alignItems: "center", 
+                  gap: "6px", 
+                  padding: "6px 28px 6px 8px", 
+                  background: "var(--md-sys-color-surface-variant)", 
+                  border: "1px solid var(--border-color)", 
+                  borderRadius: "6px", 
+                  fontSize: "0.8rem",
+                  maxWidth: "200px"
+                }}>
+                  {att.type === "image" ? (
+                    <img src={att.dataUrl} alt={att.name} style={{ width: "24px", height: "24px", objectFit: "cover", borderRadius: "3px" }} />
+                  ) : (
+                    <span style={{ fontWeight: 600 }}>📄</span>
+                  )}
+                  <span style={{ 
+                    whiteSpace: "nowrap", 
+                    overflow: "hidden", 
+                    textOverflow: "ellipsis", 
+                    color: "var(--md-sys-color-on-surface-variant)" 
+                  }} title={att.name}>
+                    {att.name}
+                  </span>
+                  <button 
+                    onClick={() => setAttachments(prev => prev.filter(item => item.id !== att.id))}
+                    style={{
+                      position: "absolute",
+                      right: "4px",
+                      top: "50%",
+                      transform: "translateY(-50%)",
+                      background: "none",
+                      border: "none",
+                      color: "var(--md-sys-color-error)",
+                      cursor: "pointer",
+                      padding: "2px",
+                      display: "flex",
+                      alignItems: "center",
+                      justifyContent: "center"
+                    }}
+                  >
+                    <X size={14} />
+                  </button>
+                </div>
+              ))}
+            </div>
+          )}
+
+          <div style={{ display: "flex", gap: "10px", width: "100%" }}>
+            <input 
+              type="file" 
+              ref={fileInputRef} 
+              style={{ display: "none" }} 
+              multiple 
+              onChange={handleFileChange} 
+            />
+            <button 
+              className="m3-btn m3-btn-tonal" 
+              onClick={() => fileInputRef.current?.click()} 
+              style={{ padding: "0 12px", height: "48px", display: "flex", alignItems: "center", justifyContent: "center" }}
+              disabled={!status.ready || isBusy}
+              title="Attach files or images"
+            >
+              <Paperclip size={17} />
+            </button>
+
+            <textarea
+              value={input}
+              onChange={(event) => setInput(event.target.value)}
+              onKeyDown={(event) => {
+                if (event.key === "Enter" && !event.shiftKey) {
+                  event.preventDefault();
+                  sendMessage();
+                }
+              }}
+              placeholder={status.ready ? "Message your local model..." : "Select and load a GGUF model above to begin"}
+              disabled={!status.ready || isBusy}
+              style={{ flex: 1 }}
+            />
+            <button 
+              className="m3-btn m3-btn-filled" 
+              onClick={sendMessage} 
+              disabled={(!input.trim() && attachments.length === 0) || !status.ready || isBusy}
+              style={{ height: "48px" }}
+            >
+              <Send size={17} /> Send
+            </button>
+          </div>
         </div>
       </section>
     </div>
diff --git a/scripts/serve.cjs b/scripts/serve.cjs
index fa2a102..e3b849d 100644
--- a/scripts/serve.cjs
+++ b/scripts/serve.cjs
@@ -220,42 +220,217 @@ function getGpuInfo() {
   if (cachedGpuInfo) return cachedGpuInfo;
 
   if (osPlatform === "win32") {
+    try {
+      const stdout = execSync(
+        'powershell -NoProfile -Command "Get-CimInstance Win32_VideoController | ForEach-Object { [PSCustomObject]@{ Name = $_.Name; VRAM = [math]::Round($_.AdapterRAM / 1GB, 2) } } | ConvertTo-Json"',
+        { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] }
+      ).trim();
+      if (stdout) {
+        let gpus = [];
+        try {
+          const parsed = JSON.parse(stdout);
+          if (Array.isArray(parsed)) {
+            gpus = parsed;
+          } else if (parsed && typeof parsed === "object") {
+            gpus = [parsed];
+          }
+        } catch (_) {}
+
+        if (gpus.length > 0) {
+          const discreteKeywords = ["nvidia", "amd", "arc", "geforce", "radeon", "rtx", "gtx"];
+          let selectedGpu = gpus.find(gpu => {
+            const name = String(gpu.Name || "").toLowerCase();
+            return discreteKeywords.some(kw => name.includes(kw));
+          });
+          if (!selectedGpu) {
+            selectedGpu = gpus[0];
+          }
+          cachedGpuInfo = {
+            name: selectedGpu.Name || "Unknown GPU",
+            vram_gb: typeof selectedGpu.VRAM === "number" ? selectedGpu.VRAM : 0
+          };
+          return cachedGpuInfo;
+        }
+      }
+    } catch (_) {}
+    // Fallback if Powershell fails or returns empty
     try {
       const output = execSync(
         "powershell -NoProfile -Command \"Get-CimInstance Win32_VideoController | Select-Object -First 1 -ExpandProperty Name\"",
         { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] },
       ).trim();
       if (output) {
-        cachedGpuInfo = { name: output };
+        cachedGpuInfo = { name: output, vram_gb: 0 };
         return cachedGpuInfo;
       }
     } catch (_) {}
   }
 
-  if (osPlatform === "linux") {
-    const linuxGpu = detectLinuxGpuFromSysfs();
-    if (linuxGpu) {
-      cachedGpuInfo = linuxGpu;
-      return cachedGpuInfo;
-    }
-  }
-
   if (osPlatform === "darwin") {
     try {
-      const output = execSync(
-        "system_profiler SPDisplaysDataType | awk -F': ' '/Chipset Model|Vendor/ {print $2; exit}'",
-        { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] },
-      ).trim();
-      if (output) {
-        cachedGpuInfo = { name: output };
-        return cachedGpuInfo;
+      const output = execSync("system_profiler SPDisplaysDataType", {
+        encoding: "utf8",
+        stdio: ["ignore", "pipe", "ignore"],
+      });
+      const lines = output.split("\n");
+      let currentGpu = null;
+      const gpus = [];
+      for (const line of lines) {
+        const trimmed = line.trim();
+        if (trimmed.startsWith("Chipset Model:")) {
+          if (currentGpu) {
+            gpus.push(currentGpu);
+          }
+          currentGpu = {
+            name: trimmed.substring("Chipset Model:".length).trim(),
+            vram_gb: 0,
+          };
+        } else if (currentGpu && (trimmed.startsWith("VRAM (Total):") || trimmed.startsWith("VRAM (Dynamic, Max):"))) {
+          const valStr = trimmed.split(":")[1].trim(); // e.g. "2 GB" or "1536 MB"
+          let bytesVal = 0;
+          const numMatch = valStr.match(/^([\d\.]+)\s*(GB|MB)/i);
+          if (numMatch) {
+            const val = parseFloat(numMatch[1]);
+            const unit = numMatch[2].toUpperCase();
+            if (unit === "GB") {
+              bytesVal = val;
+            } else if (unit === "MB") {
+              bytesVal = val / 1024;
+            }
+          }
+          currentGpu.vram_gb = Math.round(bytesVal * 100) / 100;
+        }
+      }
+      if (currentGpu) {
+        gpus.push(currentGpu);
+      }
+
+      if (gpus.length > 0) {
+        const isAppleSilicon = gpus.some(g => {
+          const name = g.name.toLowerCase();
+          return name.includes("apple") || name.includes("m1") || name.includes("m2") || name.includes("m3") || name.includes("m4");
+        }) || (os.arch() === "arm64");
+
+        if (isAppleSilicon) {
+          const totalRamGb = Math.round((os.totalmem() / (1024 ** 3)) * 100) / 100;
+          let selected = gpus.find(g => g.name.toLowerCase().includes("apple")) || gpus[0];
+          cachedGpuInfo = {
+            name: selected.name,
+            vram_gb: totalRamGb
+          };
+          return cachedGpuInfo;
+        } else {
+          const discreteKeywords = ["nvidia", "amd", "radeon", "geforce"];
+          let selected = gpus.find(gpu => {
+            const name = gpu.name.toLowerCase();
+            return discreteKeywords.some(kw => name.includes(kw));
+          });
+          if (!selected) {
+            selected = gpus[0];
+          }
+          cachedGpuInfo = {
+            name: selected.name,
+            vram_gb: selected.vram_gb
+          };
+          return cachedGpuInfo;
+        }
       }
     } catch (_) {}
-    cachedGpuInfo = { name: "Apple Metal GPU" };
+    // Fallback if system_profiler command fails
+    const totalRamGb = Math.round((os.totalmem() / (1024 ** 3)) * 100) / 100;
+    cachedGpuInfo = { name: "Apple Metal GPU", vram_gb: totalRamGb };
     return cachedGpuInfo;
   }
 
-  cachedGpuInfo = { name: "Unavailable" };
+  if (osPlatform === "linux") {
+    // AMD cards: Scan /sys/class/drm/card*/device/vendor and mem_info_vram_total
+    try {
+      if (fs.existsSync("/sys/class/drm")) {
+        const cards = fs.readdirSync("/sys/class/drm").filter(name => name.startsWith("card"));
+        for (const card of cards) {
+          const vendorPath = `/sys/class/drm/${card}/device/vendor`;
+          const vramPath = `/sys/class/drm/${card}/device/mem_info_vram_total`;
+          if (fs.existsSync(vendorPath) && fs.existsSync(vramPath)) {
+            const vendor = fs.readFileSync(vendorPath, "utf8").trim();
+            if (vendor === "0x1002") {
+              const vramBytesStr = fs.readFileSync(vramPath, "utf8").trim();
+              const vramBytes = parseFloat(vramBytesStr);
+              if (!isNaN(vramBytes)) {
+                cachedGpuInfo = {
+                  name: "AMD Radeon GPU",
+                  vram_gb: Math.round((vramBytes / (1024 ** 3)) * 100) / 100
+                };
+                return cachedGpuInfo;
+              }
+            }
+          }
+        }
+      }
+    } catch (_) {}
+
+    // NVIDIA cards: read /proc/driver/nvidia/gpus/*/information or run nvidia-smi
+    try {
+      if (fs.existsSync("/proc/driver/nvidia/gpus")) {
+        const gpuDirs = fs.readdirSync("/proc/driver/nvidia/gpus");
+        for (const gpuDir of gpuDirs) {
+          const infoPath = `/proc/driver/nvidia/gpus/${gpuDir}/information`;
+          if (fs.existsSync(infoPath)) {
+            const info = fs.readFileSync(infoPath, "utf8");
+            const modelMatch = info.match(/Model:\s+(.*)/);
+            if (modelMatch) {
+              const modelName = modelMatch[1].trim();
+              let vram_gb = 0;
+              try {
+                const smiOut = execSync("nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits", {
+                  encoding: "utf8",
+                  stdio: ["ignore", "pipe", "ignore"]
+                }).trim();
+                const parsedVal = parseFloat(smiOut);
+                if (!isNaN(parsedVal)) {
+                  vram_gb = Math.round((parsedVal / 1024) * 100) / 100;
+                }
+              } catch (_) {}
+              cachedGpuInfo = {
+                name: modelName,
+                vram_gb: vram_gb || 8.0
+              };
+              return cachedGpuInfo;
+            }
+          }
+        }
+      }
+    } catch (_) {}
+
+    try {
+      const smiOut = execSync("nvidia-smi --query-gpu=name,memory.total --format=csv,noheader,nounits", {
+        encoding: "utf8",
+        stdio: ["ignore", "pipe", "ignore"]
+      }).trim();
+      if (smiOut) {
+        const lines = smiOut.split("\n");
+        if (lines[0]) {
+          const [name, totalMb] = lines[0].split(",").map(p => p.trim());
+          const parsedVal = parseFloat(totalMb);
+          cachedGpuInfo = {
+            name: name || "NVIDIA GPU",
+            vram_gb: !isNaN(parsedVal) ? Math.round((parsedVal / 1024) * 100) / 100 : 8.0
+          };
+          return cachedGpuInfo;
+        }
+      }
+    } catch (_) {}
+
+    const linuxGpu = detectLinuxGpuFromSysfs();
+    if (linuxGpu) {
+      cachedGpuInfo = {
+        name: linuxGpu.name,
+        vram_gb: 0
+      };
+      return cachedGpuInfo;
+    }
+  }
+
+  cachedGpuInfo = { name: "Unavailable", vram_gb: 0 };
   return cachedGpuInfo;
 }
 
@@ -365,13 +540,109 @@ if (osPlatform === "darwin") {
 function getHardwareSpecs() {
   const cpus = os.cpus();
   const gpu = getGpuInfo();
+  const ramTotalGb = roundGb(os.totalmem());
+  const gpuVramGb = gpu.vram_gb || 0;
+
+  // Tiering logic
+  let tier = "low";
+  let tierReason = "";
+
+  const isAppleSilicon = osPlatform === "darwin" && (os.arch() === "arm64" || String(cpus[0]?.model || "").toLowerCase().includes("apple"));
+
+  if (isAppleSilicon) {
+    if (ramTotalGb >= 16) {
+      tier = "high";
+      tierReason = `Apple Silicon Mac with ${ramTotalGb} GB unified memory.`;
+    } else if (ramTotalGb >= 8) {
+      tier = "mid";
+      tierReason = `Apple Silicon Mac with ${ramTotalGb} GB unified memory.`;
+    } else {
+      tier = "low";
+      tierReason = `Apple Silicon Mac with ${ramTotalGb} GB unified memory.`;
+    }
+  } else {
+    if (gpuVramGb >= 12) {
+      tier = "high";
+      tierReason = `Discrete GPU with ${gpuVramGb} GB VRAM.`;
+    } else if (gpuVramGb >= 6) {
+      tier = "mid";
+      tierReason = `GPU with ${gpuVramGb} GB VRAM.`;
+    } else {
+      tier = "low";
+      tierReason = `GPU with ${gpuVramGb} GB VRAM and ${ramTotalGb} GB RAM.`;
+    }
+  }
+
+  // Recommended models catalog matching contract exactly
+  let recommended_models = [];
+  if (tier === "high") {
+    recommended_models = [
+      {
+        name: "Llama 3 8B Instruct (Q8_0)",
+        filename: "Meta-Llama-3-8B-Instruct-Q8_0.gguf",
+        approxSize: "8.5 GB",
+        url: "https://huggingface.co/QuantFactory/Meta-Llama-3-8B-Instruct-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q8_0.gguf",
+        notes: "Near-lossless precision for high quality responses.",
+        reason: tierReason
+      },
+      {
+        name: "Mistral 7B Instruct v0.3 (Q8_0)",
+        filename: "Mistral-7B-Instruct-v0.3.Q8_0.gguf",
+        approxSize: "7.7 GB",
+        url: "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.3-GGUF/resolve/main/Mistral-7B-Instruct-v0.3.Q8_0.gguf",
+        notes: "Superb instruction-following and context window size.",
+        reason: tierReason
+      }
+    ];
+  } else if (tier === "mid") {
+    recommended_models = [
+      {
+        name: "Llama 3 8B Instruct (Q4_K_M)",
+        filename: "Meta-Llama-3-8B-Instruct-Q4_K_M.gguf",
+        approxSize: "4.8 GB",
+        url: "https://huggingface.co/QuantFactory/Meta-Llama-3-8B-Instruct-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_K_M.gguf",
+        notes: "Great balance of speed and performance.",
+        reason: tierReason
+      },
+      {
+        name: "Phi-3 Mini Instruct (Q4_K_M)",
+        filename: "Phi-3-mini-4k-instruct-q4.gguf",
+        approxSize: "2.2 GB",
+        url: "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf",
+        notes: "Lightweight and fast, optimized by Microsoft.",
+        reason: tierReason
+      }
+    ];
+  } else {
+    recommended_models = [
+      {
+        name: "Qwen 2.5 Coder 0.5B Instruct (Q4_K_M)",
+        filename: "qwen2.5-coder-0.5b-instruct-q4_k_m.gguf",
+        approxSize: "491 MB",
+        url: "https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct-GGUF/resolve/main/qwen2.5-coder-0.5b-instruct-q4_k_m.gguf",
+        notes: "Extremely fast, lightweight assistant, perfect for low RAM/VRAM machines.",
+        reason: tierReason
+      },
+      {
+        name: "SmolLM2 1.7B Instruct (Q4_K_M)",
+        filename: "smollm2-1.7b-instruct-q4_k_m.gguf",
+        approxSize: "1.1 GB",
+        url: "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF/resolve/main/smollm2-1.7b-instruct-q4_k_m.gguf",
+        notes: "Excellent lightweight assistant with strong logic, reasoning, and prompt expansion capabilities.",
+        reason: tierReason
+      }
+    ];
+  }
+
   return {
     os_name: `${os.type()} ${os.release()}`,
     cpu_name: cpus[0]?.model || "Unavailable",
     cpu_cores_physical: Math.max(1, Math.round(cpus.length / 2)),
     cpu_cores_logical: cpus.length || 1,
-    ram_total_gb: roundGb(os.totalmem()),
+    ram_total_gb: ramTotalGb,
     gpu_name: gpu.name,
+    gpu_vram_gb: gpuVramGb,
+    recommended_models
   };
 }
 
@@ -381,13 +652,14 @@ function getTelemetry() {
   if (osPlatform === "darwin" && cachedMacRamUsedGb !== null) {
     ram_used_gb = cachedMacRamUsedGb;
   }
+  const gpu = getGpuInfo();
   return {
     cpu_usage: getCpuUsagePercent(),
     ram_used_gb,
     ram_total_gb: roundGb(os.totalmem()),
-    gpu_name: vram?.gpu_name || getGpuInfo().name,
+    gpu_name: vram?.gpu_name || gpu.name,
     vram_used_gb: vram?.vram_used_gb || 0,
-    vram_total_gb: vram?.vram_total_gb || 0,
+    vram_total_gb: vram?.vram_total_gb || gpu.vram_gb || 0,
   };
 }
 
@@ -1518,6 +1790,34 @@ async function startLlm(settings = {}) {
     backendBinary: path.basename(backend.path),
   };
 
+  let mmprojPath = null;
+  const lowerFilename = filename.toLowerCase();
+  const isMultimodal = lowerFilename.includes("llava") ||
+                       lowerFilename.includes("vision") ||
+                       lowerFilename.includes("qwen2vl") ||
+                       lowerFilename === "ggml-model-q4_k.gguf";
+
+  if (isMultimodal) {
+    if (settings.mmproj) {
+      const customProjPath = path.join(LLM_MODELS, path.basename(settings.mmproj));
+      if (fs.existsSync(customProjPath)) {
+        mmprojPath = customProjPath;
+      }
+    }
+    if (!mmprojPath) {
+      try {
+        const files = fs.readdirSync(LLM_MODELS);
+        const mmprojFile = files.find(file => {
+          const lower = file.toLowerCase();
+          return lower.endsWith(".gguf") && (lower === "mmproj-model-f16.gguf" || lower.includes("mmproj"));
+        });
+        if (mmprojFile) {
+          mmprojPath = path.join(LLM_MODELS, mmprojFile);
+        }
+      } catch (_) {}
+    }
+  }
+
   const args = [
     "--model", modelPath,
     "--host", "127.0.0.1",
@@ -1526,6 +1826,9 @@ async function startLlm(settings = {}) {
     "--threads", String(llmSettings.threads),
     "--n-gpu-layers", String(llmSettings.gpuLayers),
   ];
+  if (mmprojPath) {
+    args.push("--mmproj", mmprojPath);
+  }
   const spawnEnv = { ...process.env };
   const backendDir = path.dirname(backend.path);
   if (osPlatform === "linux") {
@@ -1959,7 +2262,7 @@ function describeDownloadHttpError(statusCode, url, headers = {}) {
   return `HTTP ${statusCode}`;
 }
 
-function startModelDownload(url, overrideFilename = null, targetDir = MODELS, kind = "image") {
+function startModelDownload(url, overrideFilename = null, targetDir = MODELS, kind = "image", redirectCount = 0) {
   if (downloadState.active && !overrideFilename) {
     console.log("  [download] Already downloading a model");
     return;
@@ -2035,6 +2338,10 @@ function startModelDownload(url, overrideFilename = null, targetDir = MODELS, ki
         failDownload("Redirect response did not include a Location header");
         return;
       }
+      if (redirectCount > 10) {
+        failDownload("Too many redirects");
+        return;
+      }
       console.log(`  [download] Redirected to ${redirectUrl}`);
       
       // Clean up redirected request to avoid triggering error handlers later
@@ -2045,7 +2352,7 @@ function startModelDownload(url, overrideFilename = null, targetDir = MODELS, ki
       try { fs.unlinkSync(tempPath); } catch (_) {}
       activeDownload = null;
       downloadState.active = false;
-      startModelDownload(redirectUrl, filename, targetDir, kind);
+      startModelDownload(redirectUrl, filename, targetDir, kind, redirectCount + 1);
       return;
     }