diff --git a/Cargo.lock b/Cargo.lock
index 9aa6da9c..9166f018 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1792,6 +1792,15 @@ version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
+[[package]]
+name = "libmimalloc-sys"
+version = "0.1.49"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6a45a52f43e1c16f667ccfe4dd8c85b7f7c204fd5e3bf46c5b0db9a5c3c0b8e9"
+dependencies = [
+ "cc",
+]
+
 [[package]]
 name = "libredox"
 version = "0.1.16"
@@ -1884,6 +1893,15 @@ version = "2.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f8ca58f447f06ed17d5fc4043ce1b10dd205e060fb3ce5b979b8ed8e59ff3f79"
 
+[[package]]
+name = "mimalloc"
+version = "0.1.52"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2d4139bb28d14ad1facf21d5eb8825051b326e172d216b39f6d31df53cc97862"
+dependencies = [
+ "libmimalloc-sys",
+]
+
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -1971,6 +1989,7 @@ dependencies = [
  "hex",
  "http-body-util",
  "lazy_static",
+ "mimalloc",
  "rand 0.8.6",
  "reqwest 0.12.28",
  "serde",
diff --git a/Cargo.toml b/Cargo.toml
index 73b96c64..e8a81183 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -24,6 +24,46 @@ tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 [profile.dev]
 opt-level = 3
 
+# Release profile is what the Dockerfile builds (`cargo build --release -p
+# node`). Defaults from Cargo are intentionally lax (codegen-units = 16,
+# lto = false, incremental = false), tuned for build-time and developer
+# laptops — not for a prover hot-path that runs for minutes per call.
+#
+# Three knobs are turned on here. Each is documented in
+# https://doc.rust-lang.org/cargo/reference/profiles.html.
+#
+# `lto = "thin"` enables cross-crate Link-Time Optimization on a
+# per-codegen-unit basis. Plonky2's hot path (`plonky2_field::goldilocks`
+# `Mul`, `reduce128`, the FFT inner loops) is heavily `#[inline]`d
+# intra-crate, but the `node` / `script-plonky2` boundary calls *into*
+# those generics through a trait surface. `thin` lets LLVM inline
+# across that boundary. `fat` was tried upstream and reverted (see
+# `0xPolygonZero/plonky2/Cargo.toml` — `#lto = "fat"` commented out);
+# `thin` keeps almost the full speedup with a far smaller link-time
+# regression.
+#
+# `codegen-units = 1` forces the whole crate into a single LLVM
+# compilation unit. Trades ~30-60% link-time on the node binary for
+# the most aggressive intra-crate optimization. Justified because the
+# binary's wall-clock is dominated by Plonky2 prove calls (3-15 min at
+# `MAX_IN_COINS=8`), not by `cargo build` time.
+#
+# `incremental = false` is the default for release; restating it
+# explicitly so a future `CARGO_INCREMENTAL=1` env var on a build host
+# (or a global `~/.cargo/config.toml` override) cannot silently
+# fragment the codegen unit and undo the line above. Plonky2 upstream
+# left `incremental = true` in their release profile — that is
+# explicitly counter to what we want here.
+#
+# `panic = "abort"` is *not* set: `node/src/main.rs` installs a global
+# panic hook (PR #36, MINTING_ADDRESS bootstrap fix) and the account
+# / state layers use `RwLock` poison recovery to fail a single request
+# without taking the process down. `panic = "abort"` would defeat both.
+[profile.release]
+lto = "thin"
+codegen-units = 1
+incremental = false
+
 [workspace.package]
 version = "1.1.0"
 edition = "2021"
diff --git a/node/Cargo.toml b/node/Cargo.toml
index 30cb885f..cf09cf5d 100644
--- a/node/Cargo.toml
+++ b/node/Cargo.toml
@@ -6,6 +6,21 @@ edition.workspace = true
 [dependencies]
 bitcoin = { workspace = true }
 bitcoin_hashes = { version = "0.16.0", features = ["std"] }
+# Drop-in allocator. Plonky2's prove path allocates many polynomial
+# and witness buffers across the rayon worker pool; glibc-malloc on
+# the Debian-bookworm-slim runtime image is not tuned for that
+# pattern. mimalloc consistently outperforms glibc malloc on
+# multi-thread alloc-heavy ARM64 workloads. Wired in `main.rs` via
+# `#[global_allocator]`; tests, integration tests, and the
+# `program-plonky2` / `script-plonky2` library crates keep using the
+# system allocator (no behavioural change at the prover circuit
+# level). `default-features = false` is defensive — upstream
+# (`purpleprotocol/mimalloc_rust`) currently ships `default = []`, so
+# this disables nothing today. It pins the opt-in posture for the
+# future: if upstream ever turns on `secure`, `debug`, `extended`, or
+# a v2 flip via default features, the node binary stays on the small
+# portable build until someone opts in deliberately.
+mimalloc = { version = "0.1", default-features = false }
 sha2 = { workspace = true }
 serde = { workspace = true }
 bincode = { workspace = true }
diff --git a/node/src/main.rs b/node/src/main.rs
index ead1a6fe..9c4feb5a 100644
--- a/node/src/main.rs
+++ b/node/src/main.rs
@@ -8,6 +8,16 @@
 //! duplicating definitions or making the binary itself reachable
 //! from a `cargo test --test ...` target.
 
+// mimalloc replaces the default allocator process-wide. Scoped to the
+// binary (`main.rs`) so the `node` library crate, integration tests,
+// and the `program-plonky2` / `script-plonky2` crates continue to use
+// the system allocator — keeps unit-test behaviour identical to CI
+// and avoids pulling a C build dependency into every test target.
+// See the comment on the `mimalloc` line in `node/Cargo.toml` for the
+// rationale.
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use node::account_node;
 use node::db;
 use node::publisher::EsploraConfig;