diff --git a/bench/README.md b/bench/README.md
index 548abec7..9d34fe51 100644
--- a/bench/README.md
+++ b/bench/README.md
@@ -118,24 +118,24 @@ note: encryption only affects persistence writes. GET throughput should be uncha
 ember vs chromadb vs pgvector. 100k random vectors, 128 dimensions, cosine metric, k=10 kNN search.
 HNSW index: M=16, ef_construction=64 for all systems. tested on GCP c2-standard-8.
 
-| metric | ember | chromadb | pgvector |
-|--------|-------|----------|----------|
-| insert (vectors/sec) | 917 | **3,738** | 1,562 |
-| query (queries/sec) | **1,214** | 376 | 882 |
-| query p99 (ms) | **1.09ms** | 2.90ms | 1.52ms |
-| memory (MB) | **30 MB** | 122 MB | 178 MB |
+| metric | ember | chromadb | pgvector | qdrant |
+|--------|-------|----------|----------|--------|
+| insert (vectors/sec) | 917 | **3,738** | 1,562 | — |
+| query (queries/sec) | **1,214** | 376 | 882 | — |
+| query p99 (ms) | **1.09ms** | 2.90ms | 1.52ms | — |
+| memory (MB) | **30 MB** | 122 MB | 178 MB | — |
 
 ember's query throughput is 3.2x chromadb and 1.4x pgvector, with 4-6x lower memory usage. insert throughput is lower due to per-vector RESP protocol overhead — batched pipelining helps but each VADD is still a separate command.
 
 #### SIFT1M recall accuracy (128-dim, 1M vectors, 10k queries)
 
-| metric | ember | chromadb | pgvector |
-|--------|-------|----------|----------|
-| recall@10 | — | — | — |
-| insert (vectors/sec) | — | — | — |
-| query p99 (ms) | — | — | — |
+| metric | ember | chromadb | pgvector | qdrant |
+|--------|-------|----------|----------|--------|
+| recall@10 | — | — | — | — |
+| insert (vectors/sec) | — | — | — | — |
+| query p99 (ms) | — | — | — | — |
 
-*results pending — requires a larger VM (c2-standard-16 or higher) since the 1M-vector HNSW index exceeds 16GB RAM during construction. run `bench/bench-vector.sh --sift` to populate.*
+*results pending — run `bench/bench-vector.sh --sift --qdrant` on a c2-standard-8 (32GB) to populate. SIFT1M is 1M × 128 × 4B = 512MB raw data; ember HNSW peaks around 1.5GB RSS, well within 32GB.*
 
 ### scaling efficiency
 
diff --git a/bench/bench-vector.sh b/bench/bench-vector.sh
index b4b899ec..69f65140 100755
--- a/bench/bench-vector.sh
+++ b/bench/bench-vector.sh
@@ -11,7 +11,7 @@
 #   bash bench/bench-vector.sh --ember-only # ember only (no docker needed)
 #   bash bench/bench-vector.sh --ember-grpc # include ember gRPC benchmark
 #   bash bench/bench-vector.sh --quick      # quick run (1k vectors)
-#   bash bench/bench-vector.sh --qdrant      # include qdrant comparison
+#   bash bench/bench-vector.sh --no-qdrant   # skip qdrant comparison
 #   bash bench/bench-vector.sh --sift       # SIFT1M recall accuracy
 #
 # environment variables:
@@ -46,7 +46,7 @@ BENCH_SCRIPT="bench/bench-vector.py"
 SIFT_DIR="bench/vector_data"
 
 EMBER_ONLY=false
-QDRANT=false
+QDRANT=true
 EMBER_GRPC=false
 QUICK_MODE=false
 SIFT_MODE=false
@@ -55,6 +55,7 @@ for arg in "$@"; do
     case "$arg" in
         --ember-only) EMBER_ONLY=true ;;
         --qdrant) QDRANT=true ;;
+        --no-qdrant) QDRANT=false ;;
         --ember-grpc) EMBER_GRPC=true ;;
         --quick) QUICK_MODE=true ;;
         --sift) SIFT_MODE=true ;;
diff --git a/bench/setup-vm-vector.sh b/bench/setup-vm-vector.sh
index e5d01091..e7b484fe 100755
--- a/bench/setup-vm-vector.sh
+++ b/bench/setup-vm-vector.sh
@@ -3,7 +3,7 @@
 # additional VM setup for vector similarity benchmarks.
 # run after setup-vm.sh on a fresh ubuntu VM.
 #
-# installs: python3 + deps, docker, chromadb/pgvector images
+# installs: python3 + deps, docker, chromadb/pgvector/qdrant images
 #
 # usage: ssh user@vm 'bash -s' < ./bench/setup-vm-vector.sh
 
@@ -14,7 +14,7 @@ echo "=== installing python dependencies ==="
 sudo apt-get update
 sudo apt-get install -y python3-pip python3-dev
 
-pip3 install --quiet redis chromadb psycopg2-binary numpy grpcio protobuf
+pip3 install --quiet redis chromadb psycopg2-binary numpy grpcio protobuf qdrant-client
 
 echo ""
 echo "=== installing docker ==="
@@ -33,6 +33,7 @@ echo "=== pulling docker images ==="
 
 sudo docker pull chromadb/chroma
 sudo docker pull pgvector/pgvector:pg17
+sudo docker pull qdrant/qdrant
 
 echo ""
 echo "=== installing ember python client ==="
@@ -49,8 +50,8 @@ cargo build --release -p ember-server --features jemalloc,vector,grpc
 echo ""
 echo "=== verifying ==="
 
-python3 -c "import redis, chromadb, psycopg2, numpy; print('python deps: ok')"
-sudo docker images | grep -E "chroma|pgvector" || true
+python3 -c "import redis, chromadb, psycopg2, numpy, qdrant_client; print('python deps: ok')"
+sudo docker images | grep -E "chroma|pgvector|qdrant" || true
 ./target/release/ember-server --help | head -3
 
 echo ""
@@ -61,4 +62,5 @@ echo "  ./bench/bench-vector.sh               # 100k random vectors"
 echo "  ./bench/bench-vector.sh --ember-only  # ember only (no docker)"
 echo "  ./bench/bench-vector.sh --ember-grpc  # include gRPC benchmark"
 echo "  ./bench/bench-vector.sh --quick       # quick sanity check"
+echo "  ./bench/bench-vector.sh --qdrant       # include qdrant comparison"
 echo "  ./bench/bench-vector.sh --sift        # SIFT1M recall accuracy"