diff --git a/proposals/3969-size-limits.md b/proposals/3969-size-limits.md
new file mode 100644
index 00000000000..581c798bfd4
--- /dev/null
+++ b/proposals/3969-size-limits.md
@@ -0,0 +1,260 @@
+# MSC3969: Size limits
+
+Matrix owes its name to its aim to interoperate with many protocols, some of which (Twitter,
+RFC1459 IRC, ...) have stricter message length limits than Matrix, or lower media size limit
+than typical Matrix homeservers.
+Bridges can address this by "pastebining" the text to a stable URL, and posting this URL to
+the other platform, but this can cause friction, especially when senders are unaware of
+this behavior.
+
+Additionally, room administrators may also prefer to limit the size of media and messages
+to lower resource usage of the room on participating servers, or to provide a different
+experience of Matrix.
+
+This MSC aims to provide a way for bridges and room administrators to signal size limits,
+to both media and messages should avoid using.
+This allows clients to show a warning or block users from posting messages or media which
+are larger than desired.
+
+
+## Proposal
+
+This MSC introduces a new `m.room.size_limits` state event types, which associates soft and
+hard size limits to text and media size.
+
+
+### Semantics of limits
+
+Both soft and hard limits are integers and expressed in bytes.
+
+Clients should prevent their users from exceeding the "hard" limit.
+Receiving clients and servers may hide events which do.
+If part of a message (eg. `formatted_body` or the body of a specific [MSC1767][MSC1767] block)
+exceeds a hard limit, receiving clients and servers may hide only this part (and fallback to
+either the `body` or a different [MSC1767][MSC1767] block).
+
+A "soft" limit should only show a warning in user interfaces, but clients should not prevent
+users from exceeding it. Receiving clients and servers should not hide such events because of this.
+
+If, because of wildcards, a MIME type matches the masks of multiple soft (resp. hard) limits,
+the most specific mask is used to compute the limit for that MIME type.
+If multiple masks have uncomparable specificity, the lowest limit should be considered.
+
+The number of lines includes empty lines, except the empty trailing line if any.
+(eg. `"foo\n\nbar" counts as three lines; "foobar\n" as one).
+
+If more than one `m.room.size_limits` state events are present in a room, clients should
+consider the most restrictive value of each of the limits they set.
+
+
+### Format of `m.room.size_limits`
+
+`m.room.size_limits` events may have two keys, `"soft"` and `"hard"`, each with a dictionary
+as value. These dictionaries may have the following keys:
+
+- `msgtypes` (object with positive integer values): This is a mapping from `msgtype` of `m.room.message`
+  events to the size limit of their `body` and `formatted_body`, including the `body` of
+  blocks defined by [MSC1767][MSC1767] for clients which implement it
+- `attachment_mimetypes` (object with positive integer values): This is a mapping from file mimetypes
+  to size limits. `*` may be used as a wildcard.
+- `lines` (positive integer): A limit on the number of lines in text message.
+
+Values not matching these types and undefined keys should be ignored.
+
+The state key of `m.room.size_limits` events should either follow the format defined
+in [MSC2346][MSC2346] when relevant (ie. when the limitation comes from a bridge)
+or be empty otherwise (eg. as a wish of room admins).
+
+When defining a `m.room.size_limits` with a non-empty state key, bridges should also
+publish a `uk.half-shot.bridge`/`m.bridge` event with the same key; but clients must
+be able to gracefully handle the absence of such a state event.
+
+
+### Examples
+
+
+#### RFC1459 IRC
+
+RFC1459 IRC messages are limited to 512 bytes, including message metadata overhead, whose
+length is per-user.
+IRC bridges may therefore consider the maximum possible size of the overhead, and subtract
+it from the total limit, and advertise this in `m.room.size_limits`.
+
+Sending long messages (over 3-5 lines) is usually considered poor style, but pastebining long ones
+occasionally is tolerated.
+
+For example, for an IRC server whose maximum overhead size is 100 bytes for PRIVMSG:
+
+```json5
+{
+    "content": {
+        "soft": {
+            "lines": 4
+        },
+        "hard": {
+            "msgtypes": {
+                "m.emote": 403,
+                "m.notice": 413,
+                "m.text": 412
+            }
+        }
+    },
+    "sender": "@appservice:example.com",
+    "state_key": "org.matrix.appservice-irc://irc/libera/#example",
+    "type": "m.room.size_limits"
+}
+```
+
+
+#### multiline IRC
+
+On IRC networks supporting the [`multiline`](https://ircv3.net/specs/extensions/multiline)
+capability, this is relaxed, and limits can be counted without the message overhead.
+For example, for servers advertising `max-bytes=40000,max-lines=10`, and IRC bridge could
+set this to the room:
+
+```json5
+{
+    "content": {
+        // To be nice to IRC clients on that network not implementing `multiline`:
+        "soft": {
+            "lines": 4,
+            "msgtypes": {
+                "m.emote": 403,
+                "m.notice": 413,
+                "m.text": 412
+            }
+        },
+        // The IRC server's hard limits:
+        "hard": {
+            "lines": 10,
+            "msgtypes": {
+                "m.emote": 403,
+                "m.notice": 40000,
+                "m.text": 40000
+            }
+        }
+    },
+    "sender": "@appservice:example.com",
+    "state_key": "org.matrix.appservice-irc://irc/libera/#example",
+    "type": "m.room.size_limits"
+}
+```
+
+
+#### Limiting resource use
+
+Room administrators can choose to limit resource usage for their own reasons. For example:
+
+```json5
+{
+    "content": {
+        "soft": {
+            "lines": 5,
+            "mimetypes": {
+                "text/*": 10000,  // 10kB
+                "video/*": 5000000,  // 5MB
+                "*/*": 1000000,  // 1MB for everything else
+            }
+        },
+        "hard": {
+            "lines": 100,
+            "mimetypes": {
+                "text/*": 100000,  // 100kB
+                "*/*": 1000000,  // 10MB
+            }
+        }
+    },
+    "sender": "@room-admin:example.com",
+    "state_key": "",
+    "type": "m.room.size_limits"
+}
+```
+
+## Potential issues
+
+Clients not implementing this MSC will allow their users to send events hidden from other users.
+This is already a problem when sending event types not supported by other clients, but in order
+to minimize further impact, the following precautions should be taken until this MSC is widely
+implemented:
+
+- room administrators are discouraged from configuring hard limits
+- developers are discouraged from hiding events matching either soft or hard limits
+
+Clients may disagree on line number counts from HTML messages.
+
+Bridges converting to different format to platforms with a hard limit may want to set conservative
+soft and hard limits, in order to limit the risk of users exceeding the limit after conversion.
+When this happens, they may choose to redact the offending event (to keep the room consistent with
+the bridged platform) and/or inform the user their event was dropped.
+
+Attachment limits are set by room administrators, and specific to that room.
+A homeserver's content repository operates independently of any room, so it may have different
+limits. Clients should not assume following size limits of a room implies their homeserver's
+content repository will allow their uploads.
+
+New/updated `m.room.size_limits` state events may race with new events.
+Receiving clients/servers should allow a short grace period if possible before hiding
+or dropping events.
+
+When backfilling history backward, clients should consider they may receive a
+`m.room.size_limits` event that predates events they already received.
+
+
+## Alternatives
+
+- Allowing multiple soft limits of increasing value, each with its own message/reason
+  (eg. 256 -> "this won't be bridged to Twitter", 400 -> "this is unlikely to be bridged
+  to IRC", 450 -> "this is very unlikely to be bridged to IRC", ...).
+  This could get unwieldy especially in terms of internationalization, and may not bring
+  many benefits
+- Allowing only two soft limits, with meaning "this is frowned upon" and "this is likely
+  to be dropped". It is probably better for user experience overall for "this is likely
+  to be dropped" to be the hard limit itself.
+- Dropping soft limits entirely. 
+- Making clients bundle a list of known limits for each bridge and rely on [MSC2346][MSC2346]
+  to allow them to discover bridges in a room.
+  This puts a significant maintenance burden on clients, requires users to constantly update
+  them to match new bridges (or new bridge configurations), may lead to different Matrix clients
+  showing different sets of events, does not allow room administrators to set arbitrary limits.
+- Counting grapheme clusters instead of bytes for text messages. They are more user-friendly,
+  but clients and bridged platforms implementing differing versions of Unicode would count them
+  differently.
+- Counting characters instead of bytes for text messages (like [matrix-react-sdk#4642][PR-react-sdk]).
+  Clients can easily agree on them, but they do not provide much improvement over bytes
+- Adding audio/video length limits and image/video maximum resolution. This requires at least
+  partially downloading and processing the attachment, so this brings little benefit.
+- Add a `limitation` key in (like [matrix-react-sdk#4642][PR-react-sdk]) `m.bridge` events
+  defined by [MSC2346][MSC2346] instead of a new event type. This would prevent room admins
+  from setting their own limits.
+
+
+## Security considerations
+
+Clients should not expect or rely on servers blocking or hiding events exceeding either soft
+or hard limits, even when events are not end-to-end encrypted.
+
+Different clients may see different things (but it is already an issue with `body`/`formatted_body`
+or with [MSC1767][MSC1767]), allowing abusers to send messages targeted to certain (classes of) users
+via public rooms.
+It is therefore recommended for receiving clients and servers not to hide these events from
+users with some sort of moderation powers (eg. power-level over 50, or high enough to issue
+kicks or redactions).
+
+
+## Unstable prefix
+
+During development, `org.matrix.msc3969.room.size_limits` is to be used
+instead of `m.room.size_limits`.
+
+
+## Dependencies
+
+This MSC builds on [MSC1767][MSC1767], though clients not implementing MSC1767 can simply ignore
+parts related to mimetypes, as they have nothing to enforce anyway.
+
+
+[MSC1767]: https://github.com/matrix-org/matrix-spec-proposals/pull/1767
+[MSC2346]: https://github.com/matrix-org/matrix-spec-proposals/pull/2346
+[MSC2346]: https://github.com/matrix-org/matrix-spec-proposals/pull/2346
+[PR-react-sdk]: https://github.com/matrix-org/matrix-react-sdk/pull/4642