diff --git a/README.md b/README.md
index 1c3f5f7..f845897 100644
--- a/README.md
+++ b/README.md
@@ -32,5 +32,6 @@ Runnable companion code for the Spring AI articles on [ankurm.com](https://ankur
 | [`graphrag-neo4j/`](graphrag-neo4j) | Entities and relations extracted with `ChatClient.entity(...)` into a typed Neo4j graph with a vector index on entity names (Testcontainers, `neo4j:2026.10.1-community`; also run on 5.26), then multi-hop retrieval by span linking, variable-length Cypher paths and path ranking, against plain vector RAG over the same 35 chunks and 9 questions: complete evidence 7/9 for the graph, 5/9 for vector k=8, 2/9 for k=4, with the traps reproduced (duplicate entity under an alias, an embedding merge that joins the wrong incidents, a reply with a sentence before the JSON fence, a hostile relation type, hop explosion). **The extraction step is a rule-based stand-in, not a language model**; Neo4j, embeddings and Cypher are real. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [GraphRAG in Java with Neo4j and Spring AI](https://ankurm.com/graphrag-java-neo4j-spring-ai-knowledge-graph/) |
 | [`image-gen/`](image-gen) | `ImageModel` and `ImagePrompt` with `OpenAiImageOptions` (what the real `OpenAiImageModel` puts on the wire, a per-call options object that silently resets the model, a `List<ImageMessage>` prompt that throws), a prompt gate with `ModerationModel` that fails closed, and safe storage in MinIO: sniffed content type, byte and pixel caps, server-chosen keys, a fetcher for provider URLs, presigned reads. A per-image price table from the vendors' pages dated 2026-10-28. No image provider called: the `ImageModel` and `ModerationModel` are hand-written fakes and a mock server; the MinIO tests use a real server through Testcontainers. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [Image Generation and Vision in Spring AI: ImageModel, Prompt Options and Safe Storage](https://ankurm.com/spring-ai-image-generation-imagemodel-s3-minio-safe-storage/) |
 | [`batch-llm/`](batch-llm) | 100,000 synthetic support tickets classified by a `ChatClient` inside a Spring Batch 6 chunk step: a Resilience4j rate limiter and retry around the model call (and where Spring Batch's own retry and skip belong instead), replies that are not valid JSON skipped and recorded, a restart after `kill -9` in a second JVM that ends with a byte-identical result, and a cost ledger row committed with every chunk. **The model is a deterministic simulated `ChatModel`, not a language model; token counts are estimates and prices are configurable placeholders.** Spring Boot 4.1.2, Spring Batch 6.0.6, Spring AI 2.0.2, Resilience4j 2.4.0, Java 25. | [Spring AI + Spring Batch: Classifying 100,000 Support Tickets with an LLM](https://ankurm.com/spring-ai-spring-batch-llm-ticket-classification/) |
+| [`docker-model-runner/`](docker-model-runner) | The Spring AI OpenAI starter pointed at Docker Model Runner and at Ollama by changing only `base-url` and the model name (no `/v1` is appended: the path is used exactly as given), plus a benchmark runner: time to first token, decode tokens/second, container memory from `docker stats` and tool-call success over 10 runs. Measured with the real `docker/model-runner` v1.2.9 CPU image and Ollama 0.40.13 on one shared 2-vCPU sandbox, Qwen2.5-0.5B Q4_K_M; no GPU, and Docker Desktop's `docker model` path was not used. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [Spring AI with Docker Model Runner vs Ollama: Local LLMs Benchmarked on the Same Laptop](https://ankurm.com/spring-ai-docker-model-runner-vs-ollama/) |
 
 Upgrading from Spring AI 1.x: [migration guide](https://ankurm.com/spring-ai-1-to-2-migration-guide/).
diff --git a/docker-model-runner/README.md b/docker-model-runner/README.md
new file mode 100644
index 0000000..276796b
--- /dev/null
+++ b/docker-model-runner/README.md
@@ -0,0 +1,35 @@
+# docker-model-runner
+
+Companion code for [Spring AI with Docker Model Runner vs Ollama](https://ankurm.com/spring-ai-docker-model-runner-vs-ollama/), part of the [Spring AI series](../README.md) on ankurm.com.
+
+One Spring AI `OpenAiChatModel` (the `spring-ai-starter-model-openai` starter) pointed at two local servers by changing only `spring.ai.openai.base-url` and the model name. A benchmark runner sends the same prompts to each and reports time to first token, decode tokens/second, container memory (`docker stats`) and tool-call success over N runs.
+
+Measured here: the real `docker/model-runner:latest` image (v1.2.9, llama.cpp CPU backend) and `ollama/ollama:latest` (0.40.13), both serving Qwen2.5-0.5B-Instruct at Q4_K_M, on a shared 2-vCPU sandbox with no GPU. Not measured: Docker Desktop's `docker model` CLI path and any GPU path.
+
+| Component | Version |
+|---|---|
+| Spring Boot | 4.1.2 |
+| Spring AI | 2.0.2 |
+| Java | 25 (LTS) |
+
+## Run
+
+```bash
+scripts/start-dmr.sh && scripts/start-ollama.sh   # EXTRA_CA=/path/ca.crt if HTTPS is intercepted
+# (the model sizes in run-all.sh are the 0.5B ones; pull the same models in both engines)
+scripts/run-all.sh                                # writes output/dmr.txt and output/ollama.txt
+mvn test                                          # WireTest writes output/01-wire.txt
+```
+
+On Docker Desktop with `docker model` instead: `docker model pull hf.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M`, enable host-side TCP on port 12434, then `scripts/bench.sh dmr http://localhost:12434/engines/v1 <model id from /engines/v1/models> <container name>` (memory sampling needs a container name; skip it if there is none).
+
+## Files
+
+| File | What |
+|---|---|
+| `src/main/resources/application.yml` | the whole engine configuration |
+| `Benchmark.java`, `WeatherTools.java` | throughput and tool-call harness |
+| `WireTest.java` | records what the starter sends: path, key header, model field |
+| `output/01-wire.txt` | request line, header and model name seen by a fake server |
+| `output/02-engines.txt`, `03-dmr-extra-fields.txt` | `/models` listings, image ids, usage and timing fields |
+| `output/dmr.txt`, `output/ollama.txt` | benchmark transcripts |
diff --git a/docker-model-runner/output/01-wire.txt b/docker-model-runner/output/01-wire.txt
new file mode 100644
index 0000000..de0db6d
--- /dev/null
+++ b/docker-model-runner/output/01-wire.txt
@@ -0,0 +1,8 @@
+base-url = http://127.0.0.1:<port>/engines/v1
+spring.ai.openai.chat.options.model = configured-model
+
+call 1, model.call(new Prompt("ping"))
+  POST /engines/v1/chat/completions | Authorization: Bearer not-needed | model: configured-model
+
+call 2, same prompt with OpenAiChatOptions.builder().temperature(0.0).build()
+  POST /engines/v1/chat/completions | Authorization: Bearer not-needed | model: gpt-5-mini
diff --git a/docker-model-runner/output/02-engines.txt b/docker-model-runner/output/02-engines.txt
new file mode 100644
index 0000000..de1c2b3
--- /dev/null
+++ b/docker-model-runner/output/02-engines.txt
@@ -0,0 +1,12 @@
+# GET http://localhost:12434/engines/v1/models (Docker Model Runner, docker/model-runner v1.2.9 image)
+{"object":"list","data":[{"id":"huggingface.co/qwen/qwen2.5-0.5b-instruct-gguf:Q4_K_M","object":"model","created":1726813224,"owned_by":"docker","dmr":{"context_window":32768,"architecture":"qwen2","parameters":"630.17M","quantization":"MOSTLY_Q4_K_M","size":"462.96MiB"}}]}
+
+
+# GET http://localhost:11434/v1/models (Ollama)
+{"object":"list","data":[{"id":"qwen2.5:0.5b","object":"model","created":1793253801,"owned_by":"library"},{"id":"qwen2.5:1.5b","object":"model","created":1793253450,"owned_by":"library"}]}
+
+
+# docker image digests
+docker/model-runner:latest sha256:60e6a14d8cc331fab3e51d7b78bba98274b4d28573f4e0b345e475d52d167869
+ollama/ollama:latest sha256:b2f0122c476bc9fe7ce679d8bb4bd555351e25e2554eefb156f61cb0a89cf429
+ollama version is 0.40.13
diff --git a/docker-model-runner/output/03-dmr-extra-fields.txt b/docker-model-runner/output/03-dmr-extra-fields.txt
new file mode 100644
index 0000000..eb9c652
--- /dev/null
+++ b/docker-model-runner/output/03-dmr-extra-fields.txt
@@ -0,0 +1,4 @@
+usage {'completion_tokens': 16, 'prompt_tokens': 31, 'total_tokens': 47, 'prompt_tokens_details': {'cached_tokens': 24}}
+timings keys ['cache_n', 'predicted_ms', 'predicted_n', 'predicted_per_second', 'predicted_per_token_ms', 'prompt_ms', 'prompt_n', 'prompt_per_second', 'prompt_per_token_ms']
+ollama usage {'prompt_tokens': 31, 'prompt_tokens_details': {'cached_tokens': 24}, 'completion_tokens': 10, 'total_tokens': 41}
+ollama top-level keys ['choices', 'created', 'id', 'model', 'object', 'system_fingerprint', 'usage']
diff --git a/docker-model-runner/output/dmr.txt b/docker-model-runner/output/dmr.txt
new file mode 100644
index 0000000..3e4d5cb
--- /dev/null
+++ b/docker-model-runner/output/dmr.txt
@@ -0,0 +1,56 @@
+engine label: dmr
+java: 25.0.5+8-LTS
+first request (includes model load): ttft 5018 ms, 48 completion tokens
+
+throughput, 5 prompts x 3 repeats, temperature 0, max_tokens 48
+p    rep    ttft_ms  decode_tps  wall_tps  prompt   compl
+1    1          514       19.97     16.70      41      48
+1    2           74       18.03     17.88      41      48
+1    3           91       17.50     17.26      41      48
+2    1          350       16.35     14.85      45      48
+2    2          488       13.79     12.29      45      48
+2    3           82       14.19     14.12      45      48
+3    1          315       17.07     15.60      43      48
+3    2           75       18.15     17.99      43      48
+3    3           91       17.61     17.36      43      48
+4    1          337       16.42     14.97      46      48
+4    2           87       17.94     17.70      46      48
+4    3          101       18.15     17.80      46      48
+5    1          308       17.45     15.96      45      48
+5    2           76       18.30     18.09      45      48
+5    3           89       18.15     17.88      45      48
+
+median ttft_ms 91 | median decode_tps 17.61 | median wall_tps 17.26 | n=15
+
+tool calling (one tool), 10 runs, temperature 0.7 so runs differ
+question: What is the temperature in Oslo in Celsius? Use the tools.
+expected: currentTemperature(Oslo), answer contains 17
+run  1 wrong or missing call  calls=[celsiusToFahrenheit(12)] answer=The temperature in Oslo is 54 degrees Fahrenheit.
+run  2 ok                     calls=[currentTemperature(Oslo)] answer=The temperature in Oslo is 17 degrees Celsius.
+run  3 wrong or missing call  calls=[celsiusToFahrenheit(12)] answer=The temperature in Oslo is 54 degrees Celsius.
+run  4 wrong or missing call  calls=[celsiusToFahrenheit(12)] answer=The temperature in Oslo is 54 degrees Fahrenheit.
+run  5 wrong or missing call  calls=[celsiusToFahrenheit(12)] answer=The temperature in Oslo is 54 degrees Fahrenheit.
+run  6 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo is 32 degrees Celsius.
+run  7 wrong or missing call  calls=[celsiusToFahrenheit(10)] answer=The temperature in Oslo is 50 degrees Fahrenheit.
+run  8 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo is 32 degrees Fahrenheit.
+run  9 wrong or missing call  calls=[celsiusToFahrenheit(14)] answer=The temperature in Oslo is 57 degrees Fahrenheit.
+run 10 wrong or missing call  calls=[celsiusToFahrenheit(12)] answer=The temperature in Oslo is 54 degrees Fahrenheit.
+tool-call success (one tool): 1/10 (no call 0, wrong/missing call 9, wrong answer 0, errors 0)
+
+tool calling (two tools in sequence), 10 runs, temperature 0.7 so runs differ
+question: What is the temperature in Pune in Fahrenheit? Use the tools.
+expected: currentTemperature(Pune), celsiusToFahrenheit(31), answer contains 88
+run  1 wrong or missing call  calls=[currentTemperature(Pune)] answer=The temperature in Pune is 31 degrees Fahrenheit.
+run  2 wrong or missing call  calls=[celsiusToFahrenheit(22)] answer=The temperature in Pune is 72 degrees Fahrenheit.
+run  3 wrong or missing call  calls=[currentTemperature(Pune)] answer=The temperature in Pune is 31 degrees Fahrenheit.
+run  4 wrong or missing call  calls=[celsiusToFahrenheit(25)] answer=The temperature in Pune is 77 degrees Fahrenheit.
+run  5 wrong or missing call  calls=[celsiusToFahrenheit(15)] answer=The temperature in Pune is 59 degrees Fahrenheit.
+run  6 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Pune is 32 degrees Fahrenheit.
+run  7 wrong or missing call  calls=[currentTemperature(Pune)] answer=The temperature in Pune is 31 degrees Fahrenheit.
+run  8 wrong or missing call  calls=[celsiusToFahrenheit(15)] answer=The temperature in Pune is 59 degrees Fahrenheit.
+run  9 wrong or missing call  calls=[celsiusToFahrenheit(32)] answer=The temperature in Pune is 90 degrees Fahrenheit.
+run 10 wrong or missing call  calls=[celsiusToFahrenheit(15)] answer=The temperature in Pune is 59 degrees Fahrenheit.
+tool-call success (two tools in sequence): 0/10 (no call 0, wrong/missing call 10, wrong answer 0, errors 0)
+
+memory (idle container memory before: 3.695MiB)
+container memory during run: peak 546 MiB, median 511 MiB, 58 samples
diff --git a/docker-model-runner/output/ollama.txt b/docker-model-runner/output/ollama.txt
new file mode 100644
index 0000000..dfd9b53
--- /dev/null
+++ b/docker-model-runner/output/ollama.txt
@@ -0,0 +1,56 @@
+engine label: ollama
+java: 25.0.5+8-LTS
+first request (includes model load): ttft 4946 ms, 48 completion tokens
+
+throughput, 5 prompts x 3 repeats, temperature 0, max_tokens 48
+p    rep    ttft_ms  decode_tps  wall_tps  prompt   compl
+1    1          247       17.59     16.40      41      48
+1    2           66       18.43     18.32      41      48
+1    3           63       18.02     17.93      41      48
+2    1          289       18.99     17.35      45      48
+2    2           92       18.93     18.58      45      48
+2    3           65       18.56     18.43      45      48
+3    1          322       18.36     16.63      43      48
+3    2           58       20.94     20.79      43      48
+3    3           84       17.65     17.46      43      48
+4    1          287       19.89     18.07      46      48
+4    2           68       19.30     19.15      46      48
+4    3           70       18.57     18.44      46      48
+5    1          282       17.87     16.46      45      48
+5    2           90       18.76     18.44      45      48
+5    3           68       18.74     18.55      45      48
+
+median ttft_ms 84 | median decode_tps 18.57 | median wall_tps 18.32 | n=15
+
+tool calling (one tool), 10 runs, temperature 0.7 so runs differ
+question: What is the temperature in Oslo in Celsius? Use the tools.
+expected: currentTemperature(Oslo), answer contains 17
+run  1 ok                     calls=[currentTemperature(Oslo)] answer=The current temperature in Oslo is 17 degrees Celsius.
+run  2 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo, Norway, is 32 degrees Celsius.
+run  3 wrong or missing call  calls=[celsiusToFahrenheit(20)] answer=The temperature in Oslo is 68 degrees Fahrenheit.
+run  4 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo is 32 degrees Celsius.
+run  5 wrong or missing call  calls=[celsiusToFahrenheit(10)] answer=The temperature in Oslo is currently 50 degrees Fahrenheit.
+run  6 wrong or missing call  calls=[celsiusToFahrenheit(20)] answer=The temperature in Oslo, at 20 degrees Celsius, is 68 degree
+run  7 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo in Celsius is 32 degrees.
+run  8 wrong or missing call  calls=[celsiusToFahrenheit(4)] answer=The temperature in Oslo (Oslo) is currently 39 degrees Celsi
+run  9 wrong or missing call  calls=[celsiusToFahrenheit(6)] answer=The temperature in Oslo is 43 degrees Fahrenheit.
+run 10 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Oslo is 32 degrees Celsius.
+tool-call success (one tool): 1/10 (no call 0, wrong/missing call 9, wrong answer 0, errors 0)
+
+tool calling (two tools in sequence), 10 runs, temperature 0.7 so runs differ
+question: What is the temperature in Pune in Fahrenheit? Use the tools.
+expected: currentTemperature(Pune), celsiusToFahrenheit(31), answer contains 88
+run  1 wrong or missing call  calls=[celsiusToFahrenheit(22)] answer=The temperature in Pune is 72 degrees Fahrenheit.
+run  2 wrong or missing call  calls=[celsiusToFahrenheit(32)] answer=The temperature in Pune is currently 90 degrees Fahrenheit.
+run  3 wrong or missing call  calls=[celsiusToFahrenheit(26)] answer=The temperature in Pune is 79 degrees Fahrenheit.
+run  4 wrong or missing call  calls=[celsiusToFahrenheit(25)] answer=The temperature in Pune is 77 degrees Fahrenheit.
+run  5 wrong or missing call  calls=[celsiusToFahrenheit(26)] answer=The temperature in Pune (Pune) is 79 degrees Fahrenheit. It'
+run  6 wrong or missing call  calls=[celsiusToFahrenheit(23)] answer=The temperature in Pune is 73 degrees Fahrenheit.
+run  7 wrong or missing call  calls=[celsiusToFahrenheit(23)] answer=The temperature in Pune is 73 degrees Fahrenheit.
+run  8 wrong or missing call  calls=[celsiusToFahrenheit(15)] answer=The temperature in Pune is 59 degrees Fahrenheit.
+run  9 wrong or missing call  calls=[celsiusToFahrenheit(0)] answer=The temperature in Pune is 32 degrees Fahrenheit.
+run 10 wrong or missing call  calls=[celsiusToFahrenheit(22)] answer=The temperature in Pune is 72 degrees Fahrenheit.
+tool-call success (two tools in sequence): 0/10 (no call 0, wrong/missing call 10, wrong answer 0, errors 0)
+
+memory (idle container memory before: 12.15MiB)
+container memory during run: peak 510 MiB, median 509 MiB, 55 samples
diff --git a/docker-model-runner/pom.xml b/docker-model-runner/pom.xml
new file mode 100644
index 0000000..5179ac6
--- /dev/null
+++ b/docker-model-runner/pom.xml
@@ -0,0 +1,64 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project xmlns="http://maven.apache.org/POM/4.0.0"
+         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 https://maven.apache.org/xsd/maven-4.0.0.xsd">
+  <modelVersion>4.0.0</modelVersion>
+
+  <parent>
+    <groupId>org.springframework.boot</groupId>
+    <artifactId>spring-boot-starter-parent</artifactId>
+    <version>4.1.2</version>
+    <relativePath/>
+  </parent>
+
+  <groupId>com.ankurm</groupId>
+  <artifactId>docker-model-runner</artifactId>
+  <version>1.0.0</version>
+  <name>docker-model-runner</name>
+  <description>Spring AI 2.0 against Docker Model Runner and Ollama through the same OpenAI-compatible client, with a benchmark harness for tokens per second, memory and tool-calling reliability.</description>
+
+  <properties>
+    <java.version>25</java.version>
+    <spring-ai.version>2.0.2</spring-ai.version>
+  </properties>
+
+  <dependencyManagement>
+    <dependencies>
+      <dependency>
+        <groupId>org.springframework.ai</groupId>
+        <artifactId>spring-ai-bom</artifactId>
+        <version>${spring-ai.version}</version>
+        <type>pom</type>
+        <scope>import</scope>
+      </dependency>
+    </dependencies>
+  </dependencyManagement>
+
+  <dependencies>
+    <dependency>
+      <groupId>org.springframework.ai</groupId>
+      <artifactId>spring-ai-starter-model-openai</artifactId>
+    </dependency>
+    <dependency>
+      <groupId>org.springframework.boot</groupId>
+      <artifactId>spring-boot-starter-test</artifactId>
+      <scope>test</scope>
+    </dependency>
+  </dependencies>
+
+  <build>
+    <plugins>
+      <plugin>
+        <groupId>org.springframework.boot</groupId>
+        <artifactId>spring-boot-maven-plugin</artifactId>
+      </plugin>
+      <plugin>
+        <groupId>org.apache.maven.plugins</groupId>
+        <artifactId>maven-surefire-plugin</artifactId>
+        <configuration>
+          <argLine>-Duser.timezone=UTC -Dstdout.encoding=UTF-8 -Dfile.encoding=UTF-8</argLine>
+        </configuration>
+      </plugin>
+    </plugins>
+  </build>
+</project>
diff --git a/docker-model-runner/scripts/bench.sh b/docker-model-runner/scripts/bench.sh
new file mode 100755
index 0000000..9f91017
--- /dev/null
+++ b/docker-model-runner/scripts/bench.sh
@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# bench.sh <label> <base-url> <model> <container> [extra bench args]
+# Runs the Spring AI benchmark against one OpenAI-compatible server and samples the container's
+# memory with `docker stats` once a second while it runs. Writes output/<label>.txt.
+#   Docker Model Runner : bench.sh dmr    http://localhost:12434/engines/v1 hf.co/qwen/qwen2.5-0.5b-instruct-gguf:Q4_K_M dmr
+#   Ollama              : bench.sh ollama http://localhost:11434/v1          qwen2.5:0.5b ollama
+set -euo pipefail
+cd "$(dirname "$0")/.."
+LABEL=$1; URL=$2; MODEL=$3; CTR=$4; shift 4
+[ -f target/docker-model-runner-1.0.0.jar ] || mvn -B -q -T1 package -DskipTests
+mkdir -p output
+MEM=$(mktemp)
+( while true; do docker stats --no-stream --format '{{.MemUsage}}' "$CTR" | cut -d/ -f1 >> "$MEM"; sleep 1; done ) &
+SAMPLER=$!
+trap 'kill $SAMPLER 2>/dev/null || true' EXIT
+echo "idle container memory before: $(docker stats --no-stream --format '{{.MemUsage}}' "$CTR" | cut -d/ -f1)" > "output/$LABEL.mem.tmp"
+java -jar target/docker-model-runner-1.0.0.jar --bench.run=true --bench.label="$LABEL" \
+  --bench.out-file="output/$LABEL.txt" --spring.ai.openai.base-url="$URL" \
+  --spring.ai.openai.chat.options.model="$MODEL" "$@" > "output/$LABEL.log" 2>&1 || { tail -20 "output/$LABEL.log"; exit 1; }
+kill $SAMPLER 2>/dev/null || true
+python3 - "$MEM" "output/$LABEL.mem.tmp" "output/$LABEL.txt" <<'P'
+import sys,re
+def mb(s):
+    s=s.strip(); m=re.match(r'([\d.]+)(\w+)',s)
+    if not m: return None
+    v=float(m[1]); u=m[2]
+    return v*{'B':1/1048576,'KiB':1/1024,'MiB':1,'GiB':1024}[u]
+vals=[x for x in (mb(l) for l in open(sys.argv[1])) if x is not None]
+idle=open(sys.argv[2]).read().strip()
+with open(sys.argv[3],'a') as f:
+    f.write(f"\nmemory ({idle})\ncontainer memory during run: peak {max(vals):.0f} MiB, median {sorted(vals)[len(vals)//2]:.0f} MiB, {len(vals)} samples\n")
+P
+rm -f "output/$LABEL.mem.tmp" "output/$LABEL.log"
+cat "output/$LABEL.txt"
diff --git a/docker-model-runner/scripts/run-all.sh b/docker-model-runner/scripts/run-all.sh
new file mode 100755
index 0000000..43c79bb
--- /dev/null
+++ b/docker-model-runner/scripts/run-all.sh
@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+# Regenerates output/dmr.txt and output/ollama.txt. Needs both servers up (scripts/start-dmr.sh,
+# scripts/start-ollama.sh, or your own Docker Desktop Model Runner on :12434). Each engine is
+# restarted first so its memory figure is a fresh container with only this one model loaded.
+# Run it on an otherwise idle machine; the figures are only comparable within one run.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+ARGS="${BENCH_ARGS:---bench.repeats=3 --bench.max-tokens=48 --bench.tool-runs=10}"
+docker restart ollama >/dev/null; sleep 5
+scripts/bench.sh ollama http://localhost:11434/v1 qwen2.5:0.5b ollama $ARGS
+docker restart dmr >/dev/null; sleep 5
+scripts/bench.sh dmr http://localhost:12434/engines/v1 huggingface.co/qwen/qwen2.5-0.5b-instruct-gguf:Q4_K_M dmr $ARGS
diff --git a/docker-model-runner/scripts/start-dmr.sh b/docker-model-runner/scripts/start-dmr.sh
new file mode 100755
index 0000000..c83c345
--- /dev/null
+++ b/docker-model-runner/scripts/start-dmr.sh
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Starts the real Docker Model Runner server image (docker/model-runner, llama.cpp backend, CPU
+# build here) on host port 12434 and pulls the model from Hugging Face into it.
+#
+# On a machine with Docker Desktop (4.40+) or Docker Engine plus the docker-model-plugin you do
+# NOT need this script: `docker model pull hf.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M` and
+# enable host-side TCP support (Docker Desktop: Settings > Features in development > Enable
+# host-side TCP support, port 12434). This script is the same server in a container, which is
+# what lets the benchmark run on a Linux box with no Desktop.
+#
+# EXTRA_CA=/path/to/ca.crt  (optional) bakes a corporate/proxy CA into the image, needed when
+# outbound HTTPS is intercepted: the server downloads the model itself, from inside the container.
+set -euo pipefail
+NAME="${NAME:-dmr}"
+IMAGE=docker/model-runner:latest
+if [ -n "${EXTRA_CA:-}" ]; then
+	D=$(mktemp -d); cp "$EXTRA_CA" "$D/extra-ca.crt"
+	printf 'FROM %s\nUSER root\nCOPY extra-ca.crt /usr/local/share/ca-certificates/extra-ca.crt\nRUN update-ca-certificates\nUSER modelrunner\n' "$IMAGE" > "$D/Dockerfile"
+	docker build -q -t dmr-ca "$D" >/dev/null; IMAGE=dmr-ca
+fi
+docker rm -f "$NAME" >/dev/null 2>&1 || true
+docker run -d --name "$NAME" -p 12434:12434 "$IMAGE" >/dev/null
+until curl -sf localhost:12434/engines/v1/models >/dev/null; do sleep 1; done
+curl -s -X POST localhost:12434/models/create -H 'Content-Type: application/json' \
+	-d '{"from":"hf.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M"}' | tail -c 120; echo
+curl -s localhost:12434/engines/v1/models; echo
diff --git a/docker-model-runner/scripts/start-ollama.sh b/docker-model-runner/scripts/start-ollama.sh
new file mode 100755
index 0000000..176978e
--- /dev/null
+++ b/docker-model-runner/scripts/start-ollama.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+# Starts ollama/ollama on host port 11434 and pulls qwen2.5:1.5b (Ollama's default tag is Q4_K_M,
+# the same quantisation as the GGUF the Docker Model Runner script pulls).
+# EXTRA_CA=/path/to/ca.crt  (optional) as in start-dmr.sh.
+set -euo pipefail
+NAME="${NAME:-ollama}"
+IMAGE=ollama/ollama:latest
+if [ -n "${EXTRA_CA:-}" ]; then
+	D=$(mktemp -d); cp "$EXTRA_CA" "$D/extra-ca.crt"
+	printf 'FROM %s\nCOPY extra-ca.crt /usr/local/share/ca-certificates/extra-ca.crt\nRUN update-ca-certificates\n' "$IMAGE" > "$D/Dockerfile"
+	docker build -q -t ollama-ca "$D" >/dev/null; IMAGE=ollama-ca
+fi
+docker rm -f "$NAME" >/dev/null 2>&1 || true
+docker run -d --name "$NAME" -p 11434:11434 "$IMAGE" >/dev/null
+until docker exec "$NAME" ollama list >/dev/null 2>&1; do sleep 1; done
+docker exec "$NAME" ollama pull qwen2.5:1.5b
+docker exec "$NAME" ollama show qwen2.5:1.5b | head -12
diff --git a/docker-model-runner/src/main/java/com/ankurm/dmr/BenchProperties.java b/docker-model-runner/src/main/java/com/ankurm/dmr/BenchProperties.java
new file mode 100644
index 0000000..aa3a69b
--- /dev/null
+++ b/docker-model-runner/src/main/java/com/ankurm/dmr/BenchProperties.java
@@ -0,0 +1,14 @@
+package com.ankurm.dmr;
+
+import org.springframework.boot.context.properties.ConfigurationProperties;
+
+/** Settings for the benchmark itself (the engine settings are plain spring.ai.openai.* ones). */
+@ConfigurationProperties("bench")
+public record BenchProperties(String label, Integer repeats, Integer toolRuns, Integer maxTokens, String outFile) {
+	public BenchProperties {
+		if (label == null) label = "engine";
+		if (repeats == null) repeats = 3;
+		if (toolRuns == null) toolRuns = 20;
+		if (maxTokens == null) maxTokens = 128;
+	}
+}
diff --git a/docker-model-runner/src/main/java/com/ankurm/dmr/Benchmark.java b/docker-model-runner/src/main/java/com/ankurm/dmr/Benchmark.java
new file mode 100644
index 0000000..3876164
--- /dev/null
+++ b/docker-model-runner/src/main/java/com/ankurm/dmr/Benchmark.java
@@ -0,0 +1,176 @@
+package com.ankurm.dmr;
+
+import java.io.IOException;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.time.Duration;
+import java.time.Instant;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.Locale;
+
+import org.springframework.ai.chat.client.ChatClient;
+import org.springframework.ai.chat.model.ChatModel;
+import org.springframework.ai.chat.model.ChatResponse;
+import org.springframework.ai.chat.prompt.Prompt;
+import org.springframework.ai.openai.OpenAiChatOptions;
+import org.springframework.boot.ApplicationArguments;
+import org.springframework.boot.ApplicationRunner;
+import org.springframework.boot.context.properties.EnableConfigurationProperties;
+import org.springframework.stereotype.Component;
+
+import reactor.core.publisher.Flux;
+
+/**
+ * Runs the same prompts against whichever OpenAI-compatible server spring.ai.openai.base-url
+ * points at and prints one report. Only runs when {@code --bench.run=true} is passed, so the
+ * tests can start the context without triggering it.
+ *
+ * <p>Tokens per second is computed the same way for every engine, from what the client sees:
+ * the time to the first streamed chunk, the time to the last, and the completion token count the
+ * server reports in the final usage chunk. Server-side timers are not used because Ollama's
+ * /v1 endpoint does not return them and Docker Model Runner's llama.cpp does.
+ */
+@Component
+@EnableConfigurationProperties(BenchProperties.class)
+public class Benchmark implements ApplicationRunner {
+
+	static final List<String> PROMPTS = List.of(
+			"Explain in three sentences what a Java virtual thread is.",
+			"Write a Java method that reverses a string, with a one-line comment.",
+			"List five differences between a HashMap and a TreeMap, one per line.",
+			"Summarise the CAP theorem for a new developer in under 80 words.",
+			"What does the final keyword do on a class, a method and a variable?");
+
+	private final ChatModel model;
+	private final BenchProperties props;
+
+	/**
+	 * Per-call options REPLACE the configured defaults rather than merging with them: an
+	 * OpenAiChatOptions built here without .model(..) goes out as the library default
+	 * (gpt-5-mini) and the local server answers 404. See output/01-options-replace-defaults.txt.
+	 */
+	private final String modelName;
+
+	public Benchmark(ChatModel model, BenchProperties props,
+			@org.springframework.beans.factory.annotation.Value("${spring.ai.openai.chat.options.model}") String modelName) {
+		this.model = model;
+		this.props = props;
+		this.modelName = modelName;
+	}
+
+	record Sample(double ttftMs, double decodeTps, double wallTps, int promptTokens, int completionTokens) {}
+
+	@Override
+	public void run(ApplicationArguments args) throws IOException {
+		if (!args.containsOption("bench.run")) return;
+		StringBuilder out = new StringBuilder();
+		out.append("engine label: ").append(props.label()).append('\n');
+		out.append("java: ").append(Runtime.version()).append('\n');
+
+		// One throw-away call: the first request makes the engine load the model into memory.
+		Sample cold = stream(PROMPTS.get(0));
+		out.append(String.format(Locale.ROOT, "first request (includes model load): ttft %.0f ms, %d completion tokens%n",
+				cold.ttftMs, cold.completionTokens));
+
+		List<Sample> all = new ArrayList<>();
+		out.append(String.format("%nthroughput, %d prompts x %d repeats, temperature 0, max_tokens %d%n",
+				PROMPTS.size(), props.repeats(), props.maxTokens()));
+		out.append(String.format("%-4s %-4s %9s %11s %9s %7s %7s%n", "p", "rep", "ttft_ms", "decode_tps", "wall_tps", "prompt", "compl"));
+		for (int p = 0; p < PROMPTS.size(); p++) {
+			for (int r = 1; r <= props.repeats(); r++) {
+				Sample s = stream(PROMPTS.get(p));
+				all.add(s);
+				out.append(String.format(Locale.ROOT, "%-4d %-4d %9.0f %11.2f %9.2f %7d %7d%n", p + 1, r, s.ttftMs,
+						s.decodeTps, s.wallTps, s.promptTokens, s.completionTokens));
+			}
+		}
+		out.append(String.format(Locale.ROOT, "%nmedian ttft_ms %.0f | median decode_tps %.2f | median wall_tps %.2f | n=%d%n",
+				median(all.stream().mapToDouble(Sample::ttftMs).toArray()),
+				median(all.stream().mapToDouble(Sample::decodeTps).toArray()),
+				median(all.stream().mapToDouble(Sample::wallTps).toArray()), all.size()));
+
+		out.append(toolCalling("one tool", "What is the temperature in Oslo in Celsius? Use the tools.", "currentTemperature(Oslo)", null, "17"));
+		out.append(toolCalling("two tools in sequence", "What is the temperature in Pune in Fahrenheit? Use the tools.", "currentTemperature(Pune)", "celsiusToFahrenheit(31)", "88"));
+		String report = out.toString();
+		System.out.println(report);
+		if (props.outFile() != null) Files.writeString(Path.of(props.outFile()), report);
+	}
+
+	/** One streamed completion; timings are taken on the client. */
+	Sample stream(String text) {
+		OpenAiChatOptions opts = OpenAiChatOptions.builder().model(modelName).temperature(0.0).maxTokens(props.maxTokens())
+				.streamOptions(OpenAiChatOptions.StreamOptions.builder().includeUsage(true).build()).build();
+		Instant start = Instant.now();
+		Flux<ChatResponse> flux = model.stream(new Prompt(text, opts));
+		long[] first = {0};
+		long[] last = {0};
+		int[] usage = {0, 0};
+		flux.doOnNext(r -> {
+			long now = System.nanoTime();
+			String t = r.getResult() == null || r.getResult().getOutput() == null ? null : r.getResult().getOutput().getText();
+			if (t != null && !t.isEmpty()) {
+				if (first[0] == 0) first[0] = now;
+				last[0] = now;
+			}
+			if (r.getMetadata() != null && r.getMetadata().getUsage() != null
+					&& r.getMetadata().getUsage().getCompletionTokens() != null
+					&& r.getMetadata().getUsage().getCompletionTokens() > 0) {
+				usage[0] = r.getMetadata().getUsage().getPromptTokens();
+				usage[1] = r.getMetadata().getUsage().getCompletionTokens();
+			}
+		}).blockLast(Duration.ofMinutes(10));
+		long end = System.nanoTime();
+		long begin = end - Duration.between(start, Instant.now()).toNanos();
+		double ttft = (first[0] - begin) / 1e6;
+		double decodeSecs = (last[0] - first[0]) / 1e9;
+		double decode = decodeSecs > 0 ? (usage[1] - 1) / decodeSecs : 0;
+		double wall = usage[1] / ((end - begin) / 1e9);
+		return new Sample(ttft, decode, wall, usage[0], usage[1]);
+	}
+
+	/**
+	 * N runs of a question that needs one or two tools. A run succeeds only if the expected
+	 * call(s) were made with the right arguments and the final answer contains the expected number.
+	 */
+	String toolCalling(String name, String question, String call1, String call2, String expected) {
+		int n = props.toolRuns();
+		int ok = 0, noCall = 0, wrongArgs = 0, noAnswer = 0, errors = 0;
+		StringBuilder sb = new StringBuilder();
+		sb.append(String.format("%ntool calling (%s), %d runs, temperature 0.7 so runs differ%n", name, n));
+		sb.append("question: ").append(question).append('\n');
+		sb.append("expected: ").append(call1).append(call2 == null ? "" : ", " + call2).append(", answer contains ").append(expected).append('\n');
+		for (int i = 1; i <= n; i++) {
+			WeatherTools tools = new WeatherTools();
+			String answer;
+			try {
+				answer = ChatClient.create(model).prompt().user(question)
+						.options(OpenAiChatOptions.builder().model(modelName).temperature(0.7).maxTokens(200))
+						.tools(tools).call().content();
+			} catch (RuntimeException e) {
+				errors++;
+				sb.append(String.format("run %2d ERROR %s%n", i, e.getClass().getSimpleName()));
+				continue;
+			}
+			boolean right = tools.calls.stream().anyMatch(c -> c.equalsIgnoreCase(call1))
+					&& (call2 == null || tools.calls.contains(call2));
+			boolean hasAnswer = answer != null && answer.contains(expected);
+			String verdict;
+			if (tools.calls.isEmpty()) { noCall++; verdict = "no tool call"; }
+			else if (!right) { wrongArgs++; verdict = "wrong or missing call"; }
+			else if (!hasAnswer) { noAnswer++; verdict = "answer lacks " + expected; }
+			else { ok++; verdict = "ok"; }
+			String flat = answer == null ? "null" : answer.replaceAll("\\s+", " ");
+			sb.append(String.format("run %2d %-22s calls=%s answer=%s%n", i, verdict, tools.calls, flat.substring(0, Math.min(60, flat.length()))));
+		}
+		sb.append(String.format("tool-call success (%s): %d/%d (no call %d, wrong/missing call %d, wrong answer %d, errors %d)%n",
+				name, ok, n, noCall, wrongArgs, noAnswer, errors));
+		return sb.toString();
+	}
+
+	static double median(double[] v) {
+		double[] c = v.clone();
+		java.util.Arrays.sort(c);
+		return c.length == 0 ? 0 : c.length % 2 == 1 ? c[c.length / 2] : (c[c.length / 2 - 1] + c[c.length / 2]) / 2;
+	}
+}
diff --git a/docker-model-runner/src/main/java/com/ankurm/dmr/DmrApplication.java b/docker-model-runner/src/main/java/com/ankurm/dmr/DmrApplication.java
new file mode 100644
index 0000000..683b8ea
--- /dev/null
+++ b/docker-model-runner/src/main/java/com/ankurm/dmr/DmrApplication.java
@@ -0,0 +1,16 @@
+package com.ankurm.dmr;
+
+import org.springframework.boot.SpringApplication;
+import org.springframework.boot.autoconfigure.SpringBootApplication;
+
+/**
+ * The whole application is the Spring AI OpenAI starter plus the benchmark in {@link Benchmark}.
+ * There is no Docker Model Runner (or Ollama) specific class anywhere: which engine answers is
+ * decided by {@code spring.ai.openai.base-url} and {@code spring.ai.openai.chat.options.model}.
+ */
+@SpringBootApplication
+public class DmrApplication {
+	public static void main(String[] args) {
+		SpringApplication.run(DmrApplication.class, args);
+	}
+}
diff --git a/docker-model-runner/src/main/java/com/ankurm/dmr/WeatherTools.java b/docker-model-runner/src/main/java/com/ankurm/dmr/WeatherTools.java
new file mode 100644
index 0000000..3017c81
--- /dev/null
+++ b/docker-model-runner/src/main/java/com/ankurm/dmr/WeatherTools.java
@@ -0,0 +1,34 @@
+package com.ankurm.dmr;
+
+import java.util.List;
+import java.util.concurrent.CopyOnWriteArrayList;
+
+import org.springframework.ai.tool.annotation.Tool;
+import org.springframework.ai.tool.annotation.ToolParam;
+
+/**
+ * Two deterministic tools. The temperatures are made up on purpose (31 and 17 are not what any
+ * model would guess), so a final answer containing the number proves the model read the tool
+ * result instead of answering from memory.
+ */
+public class WeatherTools {
+
+	/** Every call the model actually made, in order, as "tool(argument)". */
+	public final List<String> calls = new CopyOnWriteArrayList<>();
+
+	@Tool(description = "Get the current temperature in degrees Celsius for a city")
+	public int currentTemperature(@ToolParam(description = "City name, e.g. Pune") String city) {
+		calls.add("currentTemperature(" + city + ")");
+		return switch (city.trim().toLowerCase()) {
+			case "pune" -> 31;
+			case "oslo" -> 17;
+			default -> 22;
+		};
+	}
+
+	@Tool(description = "Convert a temperature from degrees Celsius to degrees Fahrenheit")
+	public int celsiusToFahrenheit(@ToolParam(description = "Temperature in Celsius") int celsius) {
+		calls.add("celsiusToFahrenheit(" + celsius + ")");
+		return Math.round(celsius * 9f / 5f + 32f);
+	}
+}
diff --git a/docker-model-runner/src/main/resources/application.yml b/docker-model-runner/src/main/resources/application.yml
new file mode 100644
index 0000000..da43af7
--- /dev/null
+++ b/docker-model-runner/src/main/resources/application.yml
@@ -0,0 +1,19 @@
+# The OpenAI starter used as a generic OpenAI-compatible client. Defaults point at Docker Model
+# Runner's TCP port on the host. Override on the command line to measure anything else:
+#   --spring.ai.openai.base-url=http://localhost:11434/v1 --spring.ai.openai.chat.options.model=qwen2.5:1.5b
+# base-url must END IN /v1: the starter does not append it (OpenAiSetup.OPENAI_URL is
+# "https://api.openai.com/v1") and the SDK adds only chat/completions to it.
+spring:
+  main:
+    web-application-type: none
+  ai:
+    openai:
+      base-url: http://localhost:12434/engines/v1
+      # Local engines ignore the key but the client refuses to build without one.
+      api-key: not-needed
+      # The SDK default read timeout is 60 s; a CPU-only box can need longer for one streamed answer.
+      timeout: 10m
+      chat:
+        options:
+          model: huggingface.co/qwen/qwen2.5-0.5b-instruct-gguf:Q4_K_M
+          temperature: 0.0
diff --git a/docker-model-runner/src/test/java/com/ankurm/dmr/WireTest.java b/docker-model-runner/src/test/java/com/ankurm/dmr/WireTest.java
new file mode 100644
index 0000000..fc57b17
--- /dev/null
+++ b/docker-model-runner/src/test/java/com/ankurm/dmr/WireTest.java
@@ -0,0 +1,84 @@
+package com.ankurm.dmr;
+
+import static org.assertj.core.api.Assertions.assertThat;
+
+import java.io.IOException;
+import java.net.InetSocketAddress;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.List;
+
+import com.sun.net.httpserver.HttpServer;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.BeforeEach;
+import org.junit.jupiter.api.Test;
+import org.springframework.ai.chat.model.ChatModel;
+import org.springframework.ai.chat.prompt.Prompt;
+import org.springframework.ai.openai.OpenAiChatOptions;
+import org.springframework.boot.test.context.SpringBootTest;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.test.context.DynamicPropertyRegistry;
+import org.springframework.test.context.DynamicPropertySource;
+
+/**
+ * What the OpenAI starter really puts on the wire when it is pointed at a local server. A fake
+ * server records the request line, the Authorization header and the model field, so the three
+ * questions people ask (does it add /v1? does it need a key? what model name goes out?) are
+ * answered by a request instead of by memory. Writes output/01-wire.txt.
+ */
+@SpringBootTest(properties = { "spring.ai.openai.api-key=not-needed",
+		"spring.ai.openai.chat.options.model=configured-model" })
+class WireTest {
+
+	static HttpServer server;
+	static final List<String> seen = new ArrayList<>();
+	static String prefix = "/engines/v1";
+
+	@DynamicPropertySource
+	static void props(DynamicPropertyRegistry r) throws IOException {
+		server = HttpServer.create(new InetSocketAddress("127.0.0.1", 0), 0);
+		server.createContext("/", ex -> {
+			String body = new String(ex.getRequestBody().readAllBytes(), StandardCharsets.UTF_8);
+			String model = body.replaceAll("(?s).*\"model\"\\s*:\\s*\"([^\"]*)\".*", "$1");
+			seen.add(ex.getRequestMethod() + " " + ex.getRequestURI().getPath() + " | Authorization: "
+					+ ex.getRequestHeaders().getFirst("Authorization") + " | model: " + model);
+			byte[] ok = "{\"id\":\"x\",\"object\":\"chat.completion\",\"created\":1,\"model\":\"m\",\"choices\":[{\"index\":0,\"finish_reason\":\"stop\",\"message\":{\"role\":\"assistant\",\"content\":\"pong\"}}],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":1,\"total_tokens\":4}}"
+					.getBytes(StandardCharsets.UTF_8);
+			ex.getResponseHeaders().add("Content-Type", "application/json");
+			ex.sendResponseHeaders(200, ok.length);
+			ex.getResponseBody().write(ok);
+			ex.close();
+		});
+		server.start();
+		r.add("spring.ai.openai.base-url", () -> "http://127.0.0.1:" + server.getAddress().getPort() + prefix);
+	}
+
+	@Autowired
+	ChatModel model;
+
+	@AfterEach
+	void stop() {
+		// the server is shared by the class; closed by the JVM
+	}
+
+	@Test
+	void whatGoesOnTheWire() throws IOException {
+		seen.clear();
+		model.call(new Prompt("ping"));
+		model.call(new Prompt("ping", OpenAiChatOptions.builder().temperature(0.0).build()));
+		assertThat(seen).hasSize(2);
+		// base-url is used as given: the SDK appends only chat/completions, never /v1
+		assertThat(seen.get(0)).startsWith("POST /engines/v1/chat/completions");
+		assertThat(seen.get(0)).contains("Authorization: Bearer not-needed").contains("model: configured-model");
+		// per-call options replace the configured defaults: no .model(..) means the library default
+		assertThat(seen.get(1)).doesNotContain("configured-model");
+
+		Files.createDirectories(Path.of("output"));
+		Files.writeString(Path.of("output/01-wire.txt"), "base-url = http://127.0.0.1:<port>/engines/v1\n"
+				+ "spring.ai.openai.chat.options.model = configured-model\n\n"
+				+ "call 1, model.call(new Prompt(\"ping\"))\n  " + seen.get(0) + "\n\n"
+				+ "call 2, same prompt with OpenAiChatOptions.builder().temperature(0.0).build()\n  " + seen.get(1) + "\n");
+	}
+}