offgrid-ai 0.3.17 → 0.3.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.3.17",
+  "version": "0.3.18",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/autodetect.mjs

@@ -17,9 +17,11 @@ export function detectCapabilities(modelPath, mmprojPath) {
   const nameHintsThinking = /qwen3|qwen3\.\d|gemma-4|gemma4|deepseek-r[12]/i.test(pathHints);
   const thinking = hasThinkingKwargs || nameHintsThinking;
 
-  // Quantization-aware / imatrix quantization hints. These mostly affect
-  // display and defaults transparency; llama-server does not need a QAT flag.
-  const qat = /qat|imatrix|i-?matrix/i.test(pathHints) || Object.keys(meta).some((key) => key.startsWith("quantize.imatrix."));
+  // QAT is explicit quantization-aware training lineage, mainly seen in
+  // Gemma QAT releases. imatrix is common GGUF quantization metadata and is
+  // intentionally tracked separately so we don't label every imatrix quant as QAT.
+  const qat = /\bqat\b|[-_]qat[-_]|qat[-_]?q\d/i.test(pathHints);
+  const imatrix = /imatrix|i-?matrix/i.test(pathHints) || Object.keys(meta).some((key) => key.startsWith("quantize.imatrix."));
 
   // Vision — mmproj present
   const vision = Boolean(mmprojPath && existsSync(mmprojPath));
@@ -37,7 +39,7 @@ export function detectCapabilities(modelPath, mmprojPath) {
     : undefined;
   const ctxSize = metaCtx ?? (thinking ? 80000 : 32768);
 
-  return { architecture, thinking, vision, mtp, qat, quant, metaCtx, ctxSize, meta };
+  return { architecture, thinking, vision, mtp, qat, imatrix, quant, metaCtx, ctxSize, meta };
 }
 
 // ── Compute llama-server flags from capabilities ───────────────────────────

package/src/cli.mjs

@@ -407,7 +407,8 @@ function capabilitySummary(caps) {
   if (caps.architecture) parts.push(caps.architecture);
   if (caps.quant) parts.push(caps.quant);
   if (caps.mtp) parts.push("MTP");
-  if (caps.qat) parts.push("QAT/imatrix");
+  if (caps.qat) parts.push("QAT");
+  if (caps.imatrix) parts.push("imatrix");
   if (caps.thinking) parts.push("thinking");
   if (caps.vision) parts.push("vision");
   return parts.length > 0 ? parts.join(" · ") : "standard GGUF";
@@ -417,6 +418,7 @@ function capabilityBadges(caps) {
   const badges = [];
   if (caps.mtp) badges.push(pc.blue("[MTP]"));
   if (caps.qat) badges.push(pc.green("[QAT]"));
+  if (caps.imatrix) badges.push(pc.dim("[imatrix]"));
   if (caps.thinking) badges.push(pc.magenta("[thinking]"));
   if (caps.vision) badges.push(pc.cyan("[vision]"));
   return badges.join(" ");

package/src/profile-setup.mjs

@@ -46,14 +46,23 @@ export async function configureLocalProfile(prompt, profile) {
     configured = useMtp ? applyMtpDefaults(configured) : removeMtpDefaults(configured);
   }
 
-  if (caps.thinking || caps.qat) {
+  if (caps.qat || caps.imatrix) {
     console.log("");
-    console.log(renderSection(caps.qat ? "Detected QAT / imatrix-style model" : "Detected thinking model", renderRows([
+    console.log(renderSection("Quantization note", renderRows([
+      ["QAT", caps.qat ? "yes" : "no"],
+      ["imatrix", caps.imatrix ? "yes" : "no"],
+      ["Runtime flags", "none required"],
+    ])));
+  }
+
+  if (caps.thinking) {
+    console.log("");
+    console.log(renderSection("Detected thinking model", renderRows([
       ["Defaults", "thinking / loop-safe"],
       ["Flags", "--top-k 64 --presence-penalty 0 --repeat-penalty 1.1"],
       ["Template", "--chat-template-kwargs { enable_thinking: true }"],
     ])));
-    const useThinking = await prompt.yesNo("Use these thinking/QAT-safe defaults?", true);
+    const useThinking = await prompt.yesNo("Use these thinking/loop-safe defaults?", true);
     configured = useThinking ? applyThinkingDefaults(configured) : removeThinkingDefaults(configured);
   }
 
@@ -176,7 +185,8 @@ function detectionSummary(caps) {
   if (caps.architecture) parts.push(caps.architecture);
   if (caps.quant) parts.push(caps.quant);
   if (caps.mtp) parts.push("MTP");
-  if (caps.qat) parts.push("QAT/imatrix");
+  if (caps.qat) parts.push("QAT");
+  if (caps.imatrix) parts.push("imatrix");
   if (caps.thinking) parts.push("thinking");
   if (caps.vision) parts.push("vision");
   return parts.length > 0 ? parts.join(" · ") : "standard GGUF";

package/src/profiles.mjs

@@ -165,6 +165,7 @@ function summarizeCapabilities(caps) {
     vision: caps.vision,
     mtp: caps.mtp,
     qat: caps.qat,
+    imatrix: caps.imatrix,
     quant: caps.quant,
     metaCtx: caps.metaCtx,
     ctxSize: caps.ctxSize,