diff --git a/docs/akkudoktoreos/smalldevices.md b/docs/akkudoktoreos/smalldevices.md
new file mode 100644
index 00000000..369819c5
--- /dev/null
+++ b/docs/akkudoktoreos/smalldevices.md
@@ -0,0 +1,116 @@
+% SPDX-License-Identifier: Apache-2.0
+(smalldevices-page)=
+
+# Running EOS on Small Devices
+
+EOS runs on single-board computers with 512 MB to 1 GB of RAM and on eMMC or SD storage.
+With the default configuration it needs about 290 MB at idle and 400–450 MB while a GENETIC
+run is in progress. The settings on this page bring that down to about 140 MB, without
+changing the optimization result.
+
+All figures below were measured with GENETIC at 300 individuals × 400 generations and
+15-minute slots; the individual measurements state their conditions.
+
+## Summary
+
+| Setting | Effect | Plan |
+|---|---|---|
+| `optimization.genetic.fitness_cache_max_entries: 0` [^pr1376] | −110 to −160 MB peak per run | identical |
+| `general.timezone_override` [^pr1377] | −28 MB at startup (no timezone lookup) | identical |
+| `optimization.self_consumption_interpolator: numpy` [^pr1378] | −37 MB after the first run (no SciPy) | identical |
+| `database.autosave_interval_sec: null` | storage writes from ~13 MB/min to ~1 KiB/min | identical, see notes |
+| `PYTHONMALLOC=malloc` (environment) | memory is returned after each run (glibc) | identical |
+| `OPENBLAS_NUM_THREADS=1` (environment) | no per-core BLAS buffers | identical |
+
+[^pr1376]: Added by [#1376](https://github.com/Akkudoktor-EOS/EOS/pull/1376).
+[^pr1377]: Added by [#1377](https://github.com/Akkudoktor-EOS/EOS/pull/1377), which also imports
+    optional heavy packages on first use (−123 MB at startup without any configuration change).
+[^pr1378]: Added by [#1378](https://github.com/Akkudoktor-EOS/EOS/pull/1378).
+
+Example configuration:
+
+```json
+{
+  "general": { "timezone_override": "Europe/Berlin" },
+  "optimization": {
+    "self_consumption_interpolator": "numpy",
+    "genetic": { "fitness_cache_max_entries": 0 }
+  },
+  "database": { "autosave_interval_sec": null }
+}
+```
+
+Environment of the EOS process (for example in the systemd unit):
+
+```ini
+Environment=PYTHONMALLOC=malloc
+Environment=MALLOC_ARENA_MAX=2
+Environment=MALLOC_TRIM_THRESHOLD_=131072
+Environment=MALLOC_MMAP_THRESHOLD_=131072
+Environment=OPENBLAS_NUM_THREADS=1
+```
+
+## Measurements
+
+### Memory, all settings combined
+
+Linux arm64 (Debian 13, glibc, Python 3.13), one CPU per container, server-like process
+(EOSdash off), 48 h horizon, 77 kWh battery with direct marketing; "with EV" adds an
+electric vehicle and a dynamic tariff.
+
+| | Idle | Peak during a run | Peak with EV |
+|---|---|---|---|
+| `main` (default settings) | 282–288 MB | 401 MB | 451 MB |
+| all settings above | 136 MB | 142 MB | 144 MB |
+
+With the fitness cache disabled, the memory no longer depends on the number of planned
+devices: with an EV and five home appliances the peak was 144 MB (about 300 MB with the
+unbounded cache). More devices only cost run time.
+
+### Fitness cache
+
+Same platform, two consecutive runs, `MALLOC_ARENA_MAX=2`, measured together with the lazy
+imports (server baseline 136.5 MB before the first run):
+
+| Scenario | `fitness_cache_max_entries` | Peak RSS (anon) | Run time | Best fitness |
+|---|---|---|---|---|
+| base | `null` (unbounded) | 289.3 MB (231.3) | 89.5 s | −17.473689741050 |
+| base | `0` | 179.5 MB (121.1) | 88.7 s | −17.473689741050 |
+| with EV | `null` (unbounded) | 340.0 MB (282.0) | 131.6 s | −13.851295466546 |
+| with EV | `0` | 181.3 MB (123.3) | 133.2 s | −13.851295466546 |
+
+Only 6–15 % of the cache lookups hit. A Raspberry Pi 4 with real installation data (see below)
+showed the same picture: the unbounded cache made a run about 3.5 % faster and cost 100 MB.
+
+### Run time on real hardware
+
+Data of a real installation (77 kWh battery, direct marketing, 24 h control horizon at
+15 minutes), all settings above:
+
+| Device | Run time | Peak RSS |
+|---|---|---|
+| Raspberry Pi 4 Model B (2 GB, Cortex-A72), fitness cache off | 623–636 s | 195 MB |
+| Raspberry Pi 4 Model B, fitness cache unbounded | 602–609 s | 294 MB |
+| x86 VM, 2 vCPU (container, with EV) | 173 s | 127 MB |
+
+On a Raspberry Pi 4 a run therefore takes about 10 minutes. With a 15-minute `ems.interval`
+the device is busy about 70 % of the time; a longer interval leaves more headroom.
+
+### Storage writes
+
+With the default `database.autosave_interval_sec` the measurement and prediction databases are
+written every few seconds. On an installation that pushes its forecasts and measurements from
+an external system this amounted to 12–14 MB per minute; with autosave disabled about
+1 KiB per minute remained. On SD cards and small eMMC this is the main source of wear.
+
+Disabling autosave means that data pushed to EOS is lost when EOS restarts. That is fine when an
+external system (an energy manager, Home Assistant, Node-RED) pushes forecasts and measurements
+again; it is not suitable when EOS fetches and keeps its own history.
+
+### Allocator
+
+On glibc, Python keeps memory freed after a run in its own pools. Measured on an x86 VM with an
+unbounded fitness cache, `PYTHONMALLOC=malloc` together with the trim thresholds above lowered
+the idle memory from about 295 MB to 238 MB, and after each run EOS fell back to about 260 MB
+instead of staying at its peak. With the fitness cache disabled the difference is small
+(about 5 MB), because the large per-run allocations no longer happen.
diff --git a/docs/index.md b/docs/index.md
index 10c80e05..999a0d79 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -31,6 +31,7 @@ develop/install.md
 akkudoktoreos/configuration.md
 develop/update.md
 akkudoktoreos/upgrade-genetic.md
+akkudoktoreos/smalldevices.md
 develop/revert.md
 akkudoktoreos/adapter/adapterhomeassistant.md
 akkudoktoreos/adapter/adapternodered.md