diff --git a/.forgejo/workflows/release.yml b/.forgejo/workflows/release.yml
index 1875af9..e53b875 100644
--- a/.forgejo/workflows/release.yml
+++ b/.forgejo/workflows/release.yml
@@ -16,10 +16,15 @@ jobs:
             "https://git.breadway.dev/${GITHUB_REPOSITORY}.git" src
 
       - name: build
-        run: cd src && cargo build --release --locked
+        # --features full (breadmill's npu+rocm+cuda combined) ships one binary
+        # that can use any backend at runtime via --npu/--rocm/--cuda or
+        # `backend` in config.toml. All three are ort load-dynamic (dlopen)
+        # EPs, so this doesn't require the NPU/ROCm/CUDA toolkits to be
+        # present on the build host — see breadmill/Cargo.toml.
+        run: cd src && cargo build --release --locked --workspace --features full
 
       - name: test
-        run: cd src && cargo test --release --locked --workspace
+        run: cd src && cargo test --release --locked --workspace --features full
 
       - name: prepare artifacts
         run: |
diff --git a/Cargo.lock b/Cargo.lock
index f751ae5..2e7fa2b 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -133,7 +133,7 @@ dependencies = [
 
 [[package]]
 name = "breadmill"
-version = "0.2.0"
+version = "0.2.1"
 dependencies = [
  "breadsearch-shared",
  "hex",
diff --git a/README.md b/README.md
index 700cd82..446787d 100644
--- a/README.md
+++ b/README.md
@@ -24,6 +24,7 @@ Optional features:
 | `npu`   | AMD XDNA NPU via VitisAI ONNX Runtime EP (requires Ryzen AI SDK) |
 | `rocm`  | AMD iGPU via the MIGraphX ONNX Runtime EP (ROCm-backed) |
 | `cuda`  | NVIDIA GPU via the CUDA ONNX Runtime EP |
+| `full`  | All three of the above in one binary |
 
 ```
 # NPU build
@@ -34,6 +35,9 @@ cargo build --release -p breadmill --features rocm
 
 # CUDA (NVIDIA GPU) build
 cargo build --release -p breadmill --features cuda
+
+# All backends in one binary (what the release build ships)
+cargo build --release -p breadmill --features full
 ```
 
 `rocm`/`cuda`/`npu` all use `ort`'s `load-dynamic` mode: at runtime, breadmill
@@ -43,6 +47,15 @@ build actually has the matching execution provider compiled in — breadmill
 logs a clear `Successfully registered` / `not enabled in this build` line for
 this at startup (see [GPU backend notes](#gpu-backend-notes) below).
 
+Because all three are dlopen-based, `full` doesn't require the NPU/ROCm/CUDA
+toolkits to be installed at build time — only at run time, and only for
+whichever single backend you actually select via `--npu`/`--rocm`/`--cuda`
+or `backend` in config.toml. The **released binaries are built with
+`full`**: same binary works CPU-only out of the box, and picks up NPU/ROCm/CUDA
+acceleration on a machine that has the matching ONNX Runtime available,
+without needing a different download. An explicit `--npu`/`--rocm`/`--cuda`
+flag always overrides `backend` in config.toml, not the other way around.
+
 ## Setup
 
 **1. Fetch the embedding model** (~550 MB, downloaded once from Hugging Face):
diff --git a/breadmill/Cargo.toml b/breadmill/Cargo.toml
index f9cc209..5e3822b 100644
--- a/breadmill/Cargo.toml
+++ b/breadmill/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "breadmill"
-version = "0.2.0"
+version = "0.2.1"
 edition = "2021"
 license = "MIT"
 
@@ -14,6 +14,13 @@ npu = ["ort/vitis", "ort/load-dynamic"]
 # libonnxruntime_providers_rocm.so, which most distros don't package.
 rocm = ["ort/migraphx", "ort/load-dynamic"]
 cuda = ["ort/cuda", "ort/load-dynamic"]
+# All backends in one binary. Safe to combine: every backend here uses
+# ort's load-dynamic (dlopen) mode, so none of this links against an actual
+# NPU/ROCm/CUDA toolkit at build time — which ONNX Runtime actually gets
+# loaded (and thus which EPs are really available) is decided at runtime by
+# ORT_DYLIB_PATH / the dynamic linker, per the --npu/--rocm/--cuda flag or
+# `backend` config value in use for that run.
+full = ["npu", "rocm", "cuda"]
 
 [[bin]]
 name = "breadmill"
diff --git a/breadmill/src/main.rs b/breadmill/src/main.rs
index b5bc99c..8d2e9bd 100644
--- a/breadmill/src/main.rs
+++ b/breadmill/src/main.rs
@@ -105,17 +105,34 @@ fn run_daemon(force_reindex: bool, use_npu: bool, use_rocm: bool, use_cuda: bool
     std::fs::create_dir_all(&state_dir).map_err(|e| e.to_string())?;
     std::fs::create_dir_all(&cache_dir).map_err(|e| e.to_string())?;
 
-    let backend = if use_npu || config.model.backend == "npu" {
-        eprintln!("breadmill: NPU backend selected");
-        Backend::Npu { cache_dir: cache_dir.clone() }
-    } else if use_rocm || config.model.backend == "rocm" {
-        eprintln!("breadmill: ROCm backend selected");
-        Backend::Rocm
-    } else if use_cuda || config.model.backend == "cuda" {
-        eprintln!("breadmill: CUDA backend selected");
-        Backend::Cuda
+    // CLI flags always override config — otherwise an explicit --cuda/--npu on
+    // the command line would silently lose to an unrelated `backend = "..."`
+    // already sitting in config.toml, since that's whatever earlier branch a
+    // fixed if/else-if priority order happened to check first.
+    let backend_name = if use_npu {
+        "npu"
+    } else if use_rocm {
+        "rocm"
+    } else if use_cuda {
+        "cuda"
     } else {
-        Backend::Cpu
+        config.model.backend.as_str()
+    };
+
+    let backend = match backend_name {
+        "npu" => {
+            eprintln!("breadmill: NPU backend selected");
+            Backend::Npu { cache_dir: cache_dir.clone() }
+        }
+        "rocm" => {
+            eprintln!("breadmill: ROCm backend selected");
+            Backend::Rocm
+        }
+        "cuda" => {
+            eprintln!("breadmill: CUDA backend selected");
+            Backend::Cuda
+        }
+        _ => Backend::Cpu,
     };
 
     let store = Store::open(&state_dir, dim)?;