diff --git a/src/bench-rerank.ts b/src/bench-rerank.ts
new file mode 100644
index 0000000..1887f77
--- /dev/null
+++ b/src/bench-rerank.ts
@@ -0,0 +1,319 @@
+#!/usr/bin/env bun
+/**
+ * QMD Reranker Benchmark
+ *
+ * Measures reranking performance across different configurations.
+ * Reports device, parallelism, memory, VRAM, and throughput.
+ *
+ * Usage:
+ *   bun src/bench-rerank.ts              # full benchmark
+ *   bun src/bench-rerank.ts --quick      # quick smoke test (10 docs, 1 iteration)
+ *   bun src/bench-rerank.ts --docs 100   # custom doc count
+ */
+
+import {
+  getLlama,
+  getLlamaGpuTypes,
+  resolveModelFile,
+  LlamaLogLevel,
+  type Llama,
+  type LlamaModel,
+} from "node-llama-cpp";
+import { homedir } from "os";
+import { join } from "path";
+import { cpus } from "os";
+
+// ============================================================================
+// Config
+// ============================================================================
+
+const RERANK_MODEL = "hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf";
+const MODEL_CACHE = join(homedir(), ".cache", "qmd", "models");
+const CONTEXT_SIZE = 2048;
+
+const args = process.argv.slice(2);
+const quick = args.includes("--quick");
+const docsIdx = args.indexOf("--docs");
+const DOC_COUNT = docsIdx >= 0 ? parseInt(args[docsIdx + 1]!) : (quick ? 10 : 40);
+const ITERATIONS = quick ? 1 : 3;
+const PARALLEL_CONFIGS = quick ? [1, 4] : [1, 2, 4, 8];
+
+// ============================================================================
+// Test data — realistic-ish chunks of varying length
+// ============================================================================
+
+const QUERY = "How do AI agents work and what are their limitations?";
+
+function generateDocs(n: number): string[] {
+  const templates = [
+    "Artificial intelligence agents are software systems that perceive their environment and take actions to achieve goals. They use techniques like reinforcement learning, planning, and natural language processing to operate autonomously.",
+    "The transformer architecture, introduced in 2017, revolutionized natural language processing. Self-attention mechanisms allow models to weigh the importance of different parts of input sequences when generating outputs.",
+    "Machine learning models require careful evaluation to avoid overfitting. Cross-validation, holdout sets, and metrics like precision, recall, and F1 score help assess generalization performance.",
+    "Retrieval-augmented generation combines information retrieval with language models. Documents are embedded into vector spaces, retrieved based on query similarity, and used as context for generation.",
+    "Neural network training involves forward propagation, loss computation, and backpropagation. Optimizers like Adam and SGD adjust weights to minimize the loss function over training iterations.",
+    "Large language models exhibit emergent capabilities at scale, including few-shot learning, chain-of-thought reasoning, and instruction following. These properties were not explicitly trained for.",
+    "Embedding models convert text into dense vector representations that capture semantic meaning. Similar texts produce similar vectors, enabling efficient similarity search and clustering.",
+    "Autonomous agents face challenges including hallucination, lack of grounding, limited planning horizons, and difficulty with multi-step reasoning. Safety and alignment remain open research problems.",
+    "The attention mechanism computes query-key-value interactions to determine which parts of the input are most relevant. Multi-head attention allows the model to attend to different representation subspaces.",
+    "Fine-tuning adapts a pre-trained model to specific tasks using domain-specific data. Techniques like LoRA reduce the number of trainable parameters while maintaining performance.",
+  ];
+  return Array.from({ length: n }, (_, i) => templates[i % templates.length]!);
+}
+
+// ============================================================================
+// Helpers
+// ============================================================================
+
+function formatBytes(bytes: number): string {
+  if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)} KB`;
+  if (bytes < 1024 * 1024 * 1024) return `${(bytes / (1024 * 1024)).toFixed(1)} MB`;
+  return `${(bytes / (1024 * 1024 * 1024)).toFixed(2)} GB`;
+}
+
+function getMemUsage(): { rss: number; heapUsed: number } {
+  const m = process.memoryUsage();
+  return { rss: m.rss, heapUsed: m.heapUsed };
+}
+
+function median(arr: number[]): number {
+  const sorted = [...arr].sort((a, b) => a - b);
+  const mid = Math.floor(sorted.length / 2);
+  return sorted.length % 2 !== 0 ? sorted[mid]! : (sorted[mid - 1]! + sorted[mid]!) / 2;
+}
+
+// ============================================================================
+// Benchmark runner
+// ============================================================================
+
+interface BenchResult {
+  parallelism: number;
+  contextSize: number;
+  flashAttention: boolean;
+  times: number[];       // ms per run
+  medianMs: number;
+  docsPerSec: number;
+  vramPerContext: number; // bytes
+  totalVram: number;      // bytes
+  peakRss: number;        // bytes
+}
+
+async function benchmarkConfig(
+  model: LlamaModel,
+  llama: Llama,
+  docs: string[],
+  parallelism: number,
+  flash: boolean,
+): Promise<BenchResult> {
+  // Measure VRAM before
+  const vramBefore = llama.gpu ? await llama.getVramState() : null;
+  const rssBefore = getMemUsage().rss;
+
+  // Create contexts
+  const contexts = [];
+  for (let i = 0; i < parallelism; i++) {
+    try {
+      contexts.push(await model.createRankingContext({
+        contextSize: CONTEXT_SIZE,
+        flashAttention: flash,
+      }));
+    } catch {
+      if (contexts.length === 0) {
+        // Try without flash
+        contexts.push(await model.createRankingContext({ contextSize: CONTEXT_SIZE }));
+      }
+      break;
+    }
+  }
+  const actualParallelism = contexts.length;
+
+  // Measure VRAM after context creation
+  const vramAfter = llama.gpu ? await llama.getVramState() : null;
+  const vramUsed = vramBefore && vramAfter ? vramAfter.used - vramBefore.used : 0;
+  const vramPerCtx = actualParallelism > 0 ? vramUsed / actualParallelism : 0;
+
+  // Warm up
+  await contexts[0]!.rankAll(QUERY, docs.slice(0, 2));
+
+  // Benchmark iterations
+  const times: number[] = [];
+  let peakRss = getMemUsage().rss;
+
+  for (let iter = 0; iter < ITERATIONS; iter++) {
+    const chunkSize = Math.ceil(docs.length / actualParallelism);
+
+    const t0 = performance.now();
+    const allScores = await Promise.all(
+      Array.from({ length: actualParallelism }, (_, i) => {
+        const chunk = docs.slice(i * chunkSize, (i + 1) * chunkSize);
+        return chunk.length > 0 ? contexts[i]!.rankAll(QUERY, chunk) : Promise.resolve([]);
+      })
+    );
+    const elapsed = performance.now() - t0;
+    times.push(elapsed);
+
+    // Verify scores are valid
+    const flat = allScores.flat();
+    if (flat.some(s => s < 0 || s > 1 || isNaN(s))) {
+      throw new Error("Invalid scores detected");
+    }
+
+    const currentRss = getMemUsage().rss;
+    if (currentRss > peakRss) peakRss = currentRss;
+  }
+
+  // Cleanup
+  for (const ctx of contexts) await ctx.dispose();
+
+  const med = median(times);
+  return {
+    parallelism: actualParallelism,
+    contextSize: CONTEXT_SIZE,
+    flashAttention: flash,
+    times,
+    medianMs: med,
+    docsPerSec: (docs.length / med) * 1000,
+    vramPerContext: vramPerCtx,
+    totalVram: vramUsed,
+    peakRss,
+  };
+}
+
+// ============================================================================
+// Main
+// ============================================================================
+
+async function main() {
+  console.log("═══════════════════════════════════════════════════════════════");
+  console.log("  QMD Reranker Benchmark");
+  console.log("═══════════════════════════════════════════════════════════════\n");
+
+  // Detect GPU
+  const gpuTypes = await getLlamaGpuTypes();
+  const preferred = (["cuda", "metal", "vulkan"] as const).find(g => gpuTypes.includes(g));
+
+  let llama: Llama;
+  let gpuLabel: string;
+  if (preferred) {
+    try {
+      llama = await getLlama({ gpu: preferred, logLevel: LlamaLogLevel.error });
+      gpuLabel = `${preferred}`;
+    } catch {
+      llama = await getLlama({ gpu: false, logLevel: LlamaLogLevel.error });
+      gpuLabel = "cpu (gpu init failed)";
+    }
+  } else {
+    llama = await getLlama({ gpu: false, logLevel: LlamaLogLevel.error });
+    gpuLabel = "cpu";
+  }
+
+  // System info
+  const cpuInfo = cpus();
+  const cpuModel = cpuInfo[0]?.model || "unknown";
+  const cpuCount = cpuInfo.length;
+
+  console.log("System");
+  console.log(`  CPU:       ${cpuModel}`);
+  console.log(`  Cores:     ${cpuCount} (${llama.cpuMathCores} math)`);
+  console.log(`  Device:    ${gpuLabel}`);
+
+  if (llama.gpu) {
+    const gpuNames = await llama.getGpuDeviceNames();
+    const counts = new Map<string, number>();
+    for (const name of gpuNames) counts.set(name, (counts.get(name) || 0) + 1);
+    const devStr = Array.from(counts.entries())
+      .map(([name, n]) => n > 1 ? `${n}× ${name}` : name).join(", ");
+    console.log(`  GPU:       ${devStr}`);
+    const vram = await llama.getVramState();
+    console.log(`  VRAM:      ${formatBytes(vram.total)} total, ${formatBytes(vram.free)} free`);
+  }
+
+  console.log(`  RAM:       ${formatBytes(getMemUsage().rss)} RSS at start`);
+
+  // Load model
+  console.log(`\nModel`);
+  console.log(`  URI:       ${RERANK_MODEL}`);
+  const modelPath = await resolveModelFile(RERANK_MODEL, MODEL_CACHE);
+  const vramPreModel = llama.gpu ? await llama.getVramState() : null;
+  const model = await llama.loadModel({ modelPath });
+  const vramPostModel = llama.gpu ? await llama.getVramState() : null;
+  const modelVram = vramPreModel && vramPostModel ? vramPostModel.used - vramPreModel.used : 0;
+  console.log(`  Params:    ${model.trainContextSize} train ctx`);
+  if (modelVram > 0) console.log(`  VRAM:      ${formatBytes(modelVram)} (model weights)`);
+
+  // Generate test docs
+  const docs = generateDocs(DOC_COUNT);
+  console.log(`\nBenchmark`);
+  console.log(`  Documents: ${DOC_COUNT}`);
+  console.log(`  Ctx size:  ${CONTEXT_SIZE}`);
+  console.log(`  Iterations:${ITERATIONS}`);
+  console.log(`  Query:     "${QUERY.slice(0, 50)}..."`);
+
+  // Run benchmarks
+  const results: BenchResult[] = [];
+
+  for (const p of PARALLEL_CONFIGS) {
+    if (!llama.gpu && p > 1) {
+      console.log(`\n  [${p} ctx] skipped (CPU — no benefit from parallelism)`);
+      continue;
+    }
+
+    // Test with flash attention
+    process.stdout.write(`\n  [${p} ctx, flash] running...`);
+    try {
+      const r = await benchmarkConfig(model, llama, docs, p, true);
+      results.push(r);
+      process.stdout.write(` ${r.medianMs.toFixed(0)}ms (${r.docsPerSec.toFixed(1)} docs/s)\n`);
+    } catch (e: any) {
+      process.stdout.write(` failed: ${e.message}\n`);
+      // Try without flash
+      process.stdout.write(`  [${p} ctx, no flash] running...`);
+      try {
+        const r = await benchmarkConfig(model, llama, docs, p, false);
+        results.push(r);
+        process.stdout.write(` ${r.medianMs.toFixed(0)}ms (${r.docsPerSec.toFixed(1)} docs/s)\n`);
+      } catch (e2: any) {
+        process.stdout.write(` failed: ${e2.message}\n`);
+      }
+    }
+  }
+
+  // Summary table
+  console.log("\n═══════════════════════════════════════════════════════════════");
+  console.log("  Results");
+  console.log("═══════════════════════════════════════════════════════════════\n");
+
+  const header = "  Ctx  Flash  Median    Docs/s   VRAM/ctx   Total VRAM  Peak RSS";
+  const sep    = "  ───  ─────  ──────    ──────   ────────   ──────────  ────────";
+  console.log(header);
+  console.log(sep);
+
+  const baseline = results[0]?.medianMs ?? 1;
+  for (const r of results) {
+    const speedup = baseline / r.medianMs;
+    const speedupStr = r === results[0] ? "      " : `(${speedup.toFixed(1)}×)`;
+    console.log(
+      `  ${String(r.parallelism).padStart(3)}  ` +
+      `${r.flashAttention ? " yes " : "  no "}  ` +
+      `${r.medianMs.toFixed(0).padStart(5)}ms  ` +
+      `${r.docsPerSec.toFixed(1).padStart(6)}  ` +
+      `${formatBytes(r.vramPerContext).padStart(8)}  ` +
+      `${formatBytes(r.totalVram).padStart(10)}  ` +
+      `${formatBytes(r.peakRss).padStart(8)}  ` +
+      speedupStr
+    );
+  }
+
+  // Best config
+  if (results.length > 0) {
+    const best = results.reduce((a, b) => a.docsPerSec > b.docsPerSec ? a : b);
+    console.log(`\n  Best: ${best.parallelism} contexts, flash=${best.flashAttention}`);
+    console.log(`        ${best.medianMs.toFixed(0)}ms for ${DOC_COUNT} docs (${best.docsPerSec.toFixed(1)} docs/s)`);
+    if (best.totalVram > 0) console.log(`        ${formatBytes(best.totalVram)} VRAM`);
+  }
+
+  console.log("");
+  await model.dispose();
+  await llama.dispose();
+}
+
+main().catch(console.error);