From 2272894e26398d66352f10940879d49362925857 Mon Sep 17 00:00:00 2001
From: Asher Feldman <59994+asher@users.noreply.github.com>
Date: Sun, 23 Aug 2026 08:40:56 -0700
Subject: [PATCH] chore(release): 0.4.0

---
 CHANGELOG.md           | 10 ++++++++++
 mlx_kquant/_version.py |  2 +-
 2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index ff6acfb..4606d83 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,7 +6,17 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [0.4.0]
+
 ### Changed
+- Requires mlx 0.32.1 (was 0.31.2). This is the breaking change behind
+  the major-minor bump: environments pinned to mlx 0.31.x must stay on
+  mlx-kquant 0.3.x.
+- The arena keeps a standalone Metal residency set. mlx 0.32 made its
+  queue-attached residency set private and budgeted by set_wired_limit;
+  the arena's GPU mappings now stay resident without consuming that
+  budget, so streaming-mode servers can run with the wired limit at zero.
+- Builds against nanobind 2.13.0 (was 2.12.0).
 - Quantized matmul at 4 to 8 rows is faster on machines with no NAX tile:
   an 8-row split-K tile (one simdgroup along M, four along N) replaces the
   16-row tile there, so a speculative verify of 8 rows pays for 8 rows of
diff --git a/mlx_kquant/_version.py b/mlx_kquant/_version.py
index 8a3be2e..6a9beea 100644
--- a/mlx_kquant/_version.py
+++ b/mlx_kquant/_version.py
@@ -1 +1 @@
-__version__ = "0.3.13"
+__version__ = "0.4.0"