diff --git a/speed-bench/README.md b/speed-bench/README.md
index 32075fe18..2d0844c32 100644
--- a/speed-bench/README.md
+++ b/speed-bench/README.md
@@ -17,6 +17,7 @@ Run `ds4-bench` as:
 Provide PR including your numbers if your hardware was not already tested.
 Call the benchmark csv file something like `m3_max.csv` or alike, so that
 it is clear what hardware was used for the benchmark.
+Record the machine, backend, model, and run parameters in `benchmarks.json`.
 
 To generate an SVG graph from a CSV file:
 
@@ -26,3 +27,26 @@ python3 speed-bench/plot_speed.py speed-bench/m3_max.csv --title "M3 Max t/s"
 
 The script uses only the Python standard library. By default it writes a file
 next to the CSV using the `_ts.svg` suffix, such as `speed-bench/m3_max_ts.svg`.
+
+<!-- BEGIN GENERATED BENCHMARK SUMMARY -->
+## Benchmark Summary
+
+Generated from the CSV files in this directory by `python3 speed-bench/update_summary.py`.
+
+`@ 32k ctx` means the row where `ctx_tokens` is `32768`.
+
+### DeepSeek V4 Flash q2
+
+| Hardware | Best gen (t/s) | Gen @ 32k ctx (t/s) | Avg gen (t/s) | Best prefill (t/s) | Prefill @ 32k ctx (t/s) | Avg prefill (t/s) |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Apple M4 Max | 26.8 | 24.5 | 24.6 | 344 | 248 | 250 |
+| Apple M2 Ultra | 23.2 | 21.9 | 21.9 | 411 | 326 | 325 |
+| NVIDIA DGX Spark / GB10 | 14.2 | 13.0 | 13.1 | 403 | 346 | 343 |
+
+### DeepSeek V4 PRO q2
+
+| Hardware | Best gen (t/s) | Gen @ 32k ctx (t/s) | Avg gen (t/s) | Best prefill (t/s) | Prefill @ 32k ctx (t/s) | Avg prefill (t/s) |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Apple M3 Ultra | 12.4 | 9.56 | 9.90 | 183 | 139 | 149 |
+
+<!-- END GENERATED BENCHMARK SUMMARY -->
diff --git a/speed-bench/benchmarks.json b/speed-bench/benchmarks.json
new file mode 100644
index 000000000..5f93bc2a1
--- /dev/null
+++ b/speed-bench/benchmarks.json
@@ -0,0 +1,57 @@
+{
+  "schema_version": 1,
+  "benchmarks": [
+    {
+      "csv": "gb10.csv",
+      "hardware": "NVIDIA DGX Spark / GB10",
+      "backend": "CUDA",
+      "model": "DeepSeek V4 Flash",
+      "quant": "q2",
+      "model_label": "DeepSeek V4 Flash q2",
+      "prompt_file": "speed-bench/promessi_sposi.txt",
+      "ctx_start": 2048,
+      "ctx_max": 65536,
+      "step_incr": 2048,
+      "gen_tokens": 128
+    },
+    {
+      "csv": "m2_ultra.csv",
+      "hardware": "Apple M2 Ultra",
+      "backend": "Metal",
+      "model": "DeepSeek V4 Flash",
+      "quant": "q2",
+      "model_label": "DeepSeek V4 Flash q2",
+      "prompt_file": "speed-bench/promessi_sposi.txt",
+      "ctx_start": 2048,
+      "ctx_max": 65536,
+      "step_incr": 2048,
+      "gen_tokens": 128
+    },
+    {
+      "csv": "m4_max.csv",
+      "hardware": "Apple M4 Max",
+      "backend": "Metal",
+      "model": "DeepSeek V4 Flash",
+      "quant": "q2",
+      "model_label": "DeepSeek V4 Flash q2",
+      "prompt_file": "speed-bench/promessi_sposi.txt",
+      "ctx_start": 2048,
+      "ctx_max": 65536,
+      "step_incr": 2048,
+      "gen_tokens": 128
+    },
+    {
+      "csv": "pro_model_m3_ultra.csv",
+      "hardware": "Apple M3 Ultra",
+      "backend": "Metal",
+      "model": "DeepSeek V4 PRO",
+      "quant": "q2",
+      "model_label": "DeepSeek V4 PRO q2",
+      "prompt_file": "speed-bench/promessi_sposi.txt",
+      "ctx_start": 2048,
+      "ctx_max": 32768,
+      "step_incr": 2048,
+      "gen_tokens": 128
+    }
+  ]
+}
diff --git a/speed-bench/update_summary.py b/speed-bench/update_summary.py
new file mode 100644
index 000000000..6639e5ab1
--- /dev/null
+++ b/speed-bench/update_summary.py
@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+"""Update the generated benchmark summary in speed-bench/README.md."""
+
+import csv
+import json
+from dataclasses import dataclass
+from pathlib import Path
+
+
+BEGIN_MARKER = "<!-- BEGIN GENERATED BENCHMARK SUMMARY -->"
+END_MARKER = "<!-- END GENERATED BENCHMARK SUMMARY -->"
+README = Path(__file__).with_name("README.md")
+BENCH_DIR = Path(__file__).resolve().parent
+METADATA = BENCH_DIR / "benchmarks.json"
+REQUIRED_COLUMNS = {"ctx_tokens", "prefill_tps", "gen_tps"}
+TARGET_CTX = 32768
+
+
+@dataclass(frozen=True)
+class BenchmarkMetadata:
+    csv: str
+    hardware: str
+    backend: str
+    model: str
+    quant: str
+    model_label: str
+    prompt_file: str
+    ctx_start: int
+    ctx_max: int
+    step_incr: int
+    gen_tokens: int
+
+
+@dataclass
+class BenchSummary:
+    hardware: str
+    model: str
+    best_gen: float
+    gen_at_target_ctx: float | None
+    avg_gen: float
+    best_prefill: float
+    prefill_at_target_ctx: float | None
+    avg_prefill: float
+
+
+def read_metadata() -> dict[str, BenchmarkMetadata]:
+    try:
+        data = json.loads(METADATA.read_text(encoding="utf-8"))
+    except FileNotFoundError:
+        raise SystemExit(f"{METADATA}: metadata file is required") from None
+    except json.JSONDecodeError as exc:
+        raise SystemExit(f"{METADATA}: invalid JSON: {exc}") from None
+
+    try:
+        benchmarks = data["benchmarks"]
+    except (KeyError, TypeError):
+        raise SystemExit(f"{METADATA}: expected a benchmarks list") from None
+    if not isinstance(benchmarks, list):
+        raise SystemExit(f"{METADATA}: expected a benchmarks list")
+
+    by_csv: dict[str, BenchmarkMetadata] = {}
+    for item in benchmarks:
+        try:
+            metadata = BenchmarkMetadata(**item)
+        except TypeError as exc:
+            raise SystemExit(f"{METADATA}: invalid benchmark metadata: {exc}") from None
+        if metadata.csv in by_csv:
+            raise SystemExit(f"{METADATA}: duplicate benchmark metadata for {metadata.csv}")
+        by_csv[metadata.csv] = metadata
+    return by_csv
+
+
+def fmt_tps(value: float | None) -> str:
+    if value is None:
+        return "n/a"
+    if abs(value) >= 100:
+        return f"{value:.0f}"
+    if abs(value) >= 10:
+        return f"{value:.1f}"
+    return f"{value:.2f}"
+
+
+def read_summary(path: Path, metadata: BenchmarkMetadata) -> BenchSummary:
+    rows = []
+    with path.open("r", encoding="utf-8-sig", newline="") as fp:
+        reader = csv.DictReader(fp)
+        missing = REQUIRED_COLUMNS.difference(reader.fieldnames or ())
+        if missing:
+            missing_list = ", ".join(sorted(missing))
+            raise SystemExit(f"{path}: missing CSV column(s): {missing_list}")
+
+        for row in reader:
+            rows.append(
+                {
+                    "ctx_tokens": int(row["ctx_tokens"]),
+                    "prefill_tps": float(row["prefill_tps"]),
+                    "gen_tps": float(row["gen_tps"]),
+                }
+            )
+
+    if not rows:
+        raise SystemExit(f"{path}: no benchmark rows")
+
+    target_row = next((row for row in rows if row["ctx_tokens"] == TARGET_CTX), None)
+    return BenchSummary(
+        hardware=metadata.hardware,
+        model=metadata.model_label,
+        best_gen=max(row["gen_tps"] for row in rows),
+        gen_at_target_ctx=target_row["gen_tps"] if target_row else None,
+        avg_gen=sum(row["gen_tps"] for row in rows) / len(rows),
+        best_prefill=max(row["prefill_tps"] for row in rows),
+        prefill_at_target_ctx=target_row["prefill_tps"] if target_row else None,
+        avg_prefill=sum(row["prefill_tps"] for row in rows) / len(rows),
+    )
+
+
+def render_summary(summaries: list[BenchSummary]) -> str:
+    by_model = {}
+    for summary in summaries:
+        by_model.setdefault(summary.model, []).append(summary)
+    model_groups = sorted(
+        by_model.items(),
+        key=lambda item: max(summary.best_gen for summary in item[1]),
+        reverse=True,
+    )
+    lines = [
+        BEGIN_MARKER,
+        "## Benchmark Summary",
+        "",
+        "Generated from the CSV files in this directory by `python3 speed-bench/update_summary.py`.",
+        "",
+        f"`@ 32k ctx` means the row where `ctx_tokens` is `{TARGET_CTX}`.",
+        "",
+    ]
+    for model, model_summaries in model_groups:
+        lines.extend(
+            [
+                f"### {model}",
+                "",
+                "| Hardware | Best gen (t/s) | Gen @ 32k ctx (t/s) | Avg gen (t/s) | Best prefill (t/s) | Prefill @ 32k ctx (t/s) | Avg prefill (t/s) |",
+                "| --- | ---: | ---: | ---: | ---: | ---: | ---: |",
+            ]
+        )
+        for summary in sorted(model_summaries, key=lambda item: item.best_gen, reverse=True):
+            lines.append(
+                "| "
+                + " | ".join(
+                    [
+                        summary.hardware,
+                        fmt_tps(summary.best_gen),
+                        fmt_tps(summary.gen_at_target_ctx),
+                        fmt_tps(summary.avg_gen),
+                        fmt_tps(summary.best_prefill),
+                        fmt_tps(summary.prefill_at_target_ctx),
+                        fmt_tps(summary.avg_prefill),
+                    ]
+                )
+                + " |"
+            )
+        lines.append("")
+    lines.extend([END_MARKER, ""])
+    return "\n".join(lines)
+
+
+def replace_generated_section(readme: str, generated: str) -> str:
+    begin = readme.find(BEGIN_MARKER)
+    end = readme.find(END_MARKER)
+    if begin == -1 and end == -1:
+        return readme.rstrip() + "\n\n" + generated
+    if begin == -1 or end == -1 or end < begin:
+        raise SystemExit("README.md has mismatched generated summary markers")
+    end += len(END_MARKER)
+    return readme[:begin].rstrip() + "\n\n" + generated.rstrip() + readme[end:].rstrip() + "\n"
+
+
+def main() -> None:
+    csv_paths = sorted(BENCH_DIR.glob("*.csv"))
+    if not csv_paths:
+        raise SystemExit(f"{BENCH_DIR}: no CSV files found")
+    metadata = read_metadata()
+    missing = [path.name for path in csv_paths if path.name not in metadata]
+    if missing:
+        missing_list = ", ".join(missing)
+        raise SystemExit(f"{METADATA}: missing metadata for CSV file(s): {missing_list}")
+    summaries = [read_summary(path, metadata[path.name]) for path in csv_paths]
+    generated = render_summary(summaries)
+    README.write_text(replace_generated_section(README.read_text(encoding="utf-8"), generated), encoding="utf-8")
+
+
+if __name__ == "__main__":
+    main()