diff --git a/cake-cli/src/main.rs b/cake-cli/src/main.rs
index 7bc79a3..3c99c74 100644
--- a/cake-cli/src/main.rs
+++ b/cake-cli/src/main.rs
@@ -308,6 +308,6 @@ fn build_worker_topology(
 /// Return the base cache directory for zero-config model data.
 fn cache_base_dir() -> PathBuf {
     dirs::cache_dir()
-        .unwrap_or_else(|| PathBuf::from("/tmp"))
+        .unwrap_or_else(std::env::temp_dir)
         .join("cake")
 }
diff --git a/cake-core/src/cake/discovery.rs b/cake-core/src/cake/discovery.rs
index 19c3440..b5c3ac5 100644
--- a/cake-core/src/cake/discovery.rs
+++ b/cake-core/src/cake/discovery.rs
@@ -243,10 +243,26 @@ pub fn detect_backend() -> String {
 
 /// Detect the CUDA toolkit version via nvcc.
 pub fn detect_cuda_version() -> Option<String> {
+    // Try nvcc from PATH first.
     let output = std::process::Command::new("nvcc")
         .arg("--version")
-        .output()
-        .ok()?;
+        .output();
+
+    // On Windows, nvcc may not be in PATH but CUDA_PATH is always set by the installer.
+    // Only fall back to CUDA_PATH if the PATH lookup failed.
+    #[cfg(target_os = "windows")]
+    let output = output.or_else(|_| {
+        let cuda_path = std::env::var("CUDA_PATH")
+            .map_err(|_| std::io::Error::new(std::io::ErrorKind::NotFound, "CUDA_PATH not set"))?;
+        let nvcc_path = std::path::PathBuf::from(&cuda_path)
+            .join("bin")
+            .join("nvcc.exe");
+        std::process::Command::new(nvcc_path)
+            .arg("--version")
+            .output()
+    });
+
+    let output = output.ok()?;
     if !output.status.success() {
         return None;
     }
@@ -330,6 +346,24 @@ fn detect_system_memory() -> u64 {
         }
     }
 
+    // On Windows, get total physical memory.
+    // Try PowerShell/CIM first (wmic is deprecated and removed in some Win11 builds).
+    #[cfg(target_os = "windows")]
+    {
+        if let Ok(output) = std::process::Command::new("powershell")
+            .args(["-NoProfile", "-Command",
+                   "(Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory"])
+            .output()
+        {
+            if output.status.success() {
+                let stdout = String::from_utf8_lossy(&output.stdout);
+                if let Ok(bytes) = stdout.trim().parse::<u64>() {
+                    return bytes;
+                }
+            }
+        }
+    }
+
     // On Linux and Android, read /proc/meminfo
     // Note: Android uses target_os = "android", not "linux", so both are listed.
     #[cfg(any(target_os = "linux", target_os = "android"))]
@@ -530,6 +564,52 @@ fn get_broadcast_addresses() -> Vec<Ipv4Addr> {
         }
     }
 
+    // On Windows, parse `ipconfig` to compute broadcast addresses from IP + subnet mask.
+    // ipconfig groups lines under adapter headers (non-indented). Indented lines contain
+    // the IPv4 address and subnet mask. We reset last_ip on adapter boundaries so a
+    // stale IP from a previous adapter can't pair with the wrong subnet mask.
+    #[cfg(target_os = "windows")]
+    {
+        if let Ok(output) = std::process::Command::new("ipconfig").output() {
+            let stdout = String::from_utf8_lossy(&output.stdout);
+            let mut last_ip: Option<Ipv4Addr> = None;
+            for line in stdout.lines() {
+                // Adapter headers have no leading whitespace; detail lines are indented.
+                // Reset state on adapter boundaries to prevent cross-adapter mispairing.
+                if !line.starts_with(' ') && !line.starts_with('\t') {
+                    last_ip = None;
+                }
+
+                let trimmed = line.trim();
+                // Match lines containing an IPv4 address value (x.x.x.x after a colon).
+                // This is locale-independent: we look for ": <ipv4>" on any line and
+                // distinguish address vs mask by checking if it looks like a mask
+                // (starts with 255.).
+                if let Some(colon_idx) = trimmed.rfind(':') {
+                    let value = trimmed[colon_idx + 1..].trim();
+                    if let Ok(ip) = value.parse::<Ipv4Addr>() {
+                        let octets = ip.octets();
+                        if octets[0] == 255 {
+                            // This is a subnet mask — pair with last_ip
+                            if let Some(addr) = last_ip.take() {
+                                let ip_bits = u32::from(addr);
+                                let mask_bits = u32::from(ip);
+                                let brd = (ip_bits & mask_bits) | (!mask_bits);
+                                let brd_addr = Ipv4Addr::from(brd);
+                                if !brd_addr.is_loopback() {
+                                    addrs.push(brd_addr);
+                                }
+                            }
+                        } else if !ip.is_loopback() {
+                            // This is an IPv4 address
+                            last_ip = Some(ip);
+                        }
+                    }
+                }
+            }
+        }
+    }
+
     // Always include the limited broadcast as a fallback
     addrs.push(Ipv4Addr::BROADCAST);
     addrs.dedup();
diff --git a/cake-core/src/utils/models.rs b/cake-core/src/utils/models.rs
index 4d4245d..d84b22f 100644
--- a/cake-core/src/utils/models.rs
+++ b/cake-core/src/utils/models.rs
@@ -171,7 +171,7 @@ fn hf_cache_dir() -> Option<PathBuf> {
 /// Return the Cake cluster cache directory if it exists.
 fn cake_cache_dir() -> Option<PathBuf> {
     let cache = dirs::cache_dir()
-        .unwrap_or_else(|| PathBuf::from("/tmp"))
+        .unwrap_or_else(std::env::temp_dir)
         .join("cake");
     if cache.exists() {
         Some(cache)
diff --git a/docs/install.md b/docs/install.md
index cdb12a4..9af472a 100644
--- a/docs/install.md
+++ b/docs/install.md
@@ -48,6 +48,14 @@ version supported by your driver to avoid library version mismatches:
 CUDA_HOME=/usr/local/cuda-12.4 cargo build --release --features cuda
 ```
 
+**CUDA on Windows:**
+
+Windows workers require an NVIDIA GPU driver and the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) >= 12.2 (the installer sets `CUDA_PATH` automatically).
+
+```powershell
+cargo build --release --features cuda
+```
+
 ### Pre-Volta NVIDIA GPUs
 
 For Pascal, Maxwell, or other GPUs with compute capability < 7.0, the upstream `candle-kernels` crate requires patches. See [`cuda-compat/`](../cuda-compat/) for a one-command fix: