diff --git a/zig/build/cuda.zig b/zig/build/cuda.zig
index 5fadb17f06..2573889f3a 100644
--- a/zig/build/cuda.zig
+++ b/zig/build/cuda.zig
@@ -225,7 +225,8 @@ pub fn hostTests(b: *std.Build, draft_ids: *std.Build.Module, build_options: *st
     const host = b.graph.host;
     step.dependOn(&b.addRunArtifact(b.addTest(.{ .root_module = b.createModule(.{ .root_source_file = b.path("zig/build/check_fatbins.zig"), .target = host }) })).step);
     const cuda = runtime(b, host, .debug, &.{});
-    @import("h3_host.zig").targets(b, cuda, step);
+    // H3's library opens its checkpoint with mmap: not Windows
+    if (host.result.os.tag != .windows) @import("h3_host.zig").targets(b, cuda, step);
     const mods = family(b, host, .debug, cuda, draft_ids);
     @import("qwen_contract.zig").host(b, mods.core, step);
     const mtp_policy = b.createModule(.{ .root_source_file = b.path("zig/nemotron_mtp_policy_host.zig"), .target = host, .optimize = .debug, .link_libc = true, .imports = &.{ .{ .name = "nemotron", .module = mods.nemotron }, .{ .name = "lanes", .module = mods.lanes } } });
@@ -358,7 +359,8 @@ pub fn hostTests(b: *std.Build, draft_ids: *std.Build.Module, build_options: *st
     _ = open_object.getEmittedBin();
     b.step("compile-omni-video-host", "Compile actual native CUDA startup and offer for Linux without images or GPU execution").dependOn(&open_object.step);
     const h3_checkpoint = b.createModule(.{ .root_source_file = b.path("zig/src/families/h3/checkpoint.zig"), .target = host, .optimize = .debug, .link_libc = true });
-    step.dependOn(&b.addRunArtifact(b.addTest(.{ .root_module = h3_checkpoint })).step);
+    // the checkpoint maps its file with mmap: not Windows
+    if (host.result.os.tag != .windows) step.dependOn(&b.addRunArtifact(b.addTest(.{ .root_module = h3_checkpoint })).step);
     // the HTTP server's unit tests (zig/src/server/root.zig), which the macOS build runs, on Linux hosts too
     const template = b.createModule(.{ .root_source_file = b.path("zig/src/core/template/template.zig"), .target = host, .optimize = .debug, .link_libc = true });
     const server_tests = b.createModule(.{
diff --git a/zig/src/core/checkpoint_identity_test.zig b/zig/src/core/checkpoint_identity_test.zig
index a058bca520..c2e3857c8a 100644
--- a/zig/src/core/checkpoint_identity_test.zig
+++ b/zig/src/core/checkpoint_identity_test.zig
@@ -124,6 +124,7 @@ fn fixturePaths(a: std.mem.Allocator, tmp: *const std.testing.TmpDir, sub: []con
 }
 
 test "whole and split consumers share cached file digests and still check manifests" {
+    if (!@import("digest_key.zig").persistent) return error.SkipZigTest;
     const a = std.testing.allocator;
     const io = std.testing.io;
     var tmp = std.testing.tmpDir(.{});
@@ -168,6 +169,7 @@ test "full checkpoint identity includes payloads and detects changed files at re
 }
 
 test "a corrupt shared digest cache refuses the real content consumer" {
+    if (!@import("digest_key.zig").persistent) return error.SkipZigTest;
     const a = std.testing.allocator;
     const io = std.testing.io;
     var tmp = std.testing.tmpDir(.{});
diff --git a/zig/src/core/digest_cache_probe.zig b/zig/src/core/digest_cache_probe.zig
index ad75b8c90e..87ddf5e253 100644
--- a/zig/src/core/digest_cache_probe.zig
+++ b/zig/src/core/digest_cache_probe.zig
@@ -30,5 +30,7 @@ pub fn main(init: std.process.Init) !void {
     }
     var output: [32]u8 = undefined;
     const text = try std.fmt.bufPrint(&output, "{d}\n", .{calls});
+    // std.c.write takes a HANDLE on Windows; this executable is built without the sys module
+    if (@import("builtin").os.tag == .windows) return std.Io.File.stdout().writeStreamingAll(init.io, text);
     if (std.c.write(1, text.ptr, text.len) != text.len) return error.Output;
 }
diff --git a/zig/src/core/digest_key.zig b/zig/src/core/digest_key.zig
index 966f1fd223..e3682575ee 100644
--- a/zig/src/core/digest_key.zig
+++ b/zig/src/core/digest_key.zig
@@ -31,7 +31,7 @@ pub const Key = struct {
         }
         const s = try file.stat(io);
         if (s.kind != .file) return error.DigestNotRegular;
-        return .{ .device = 0, .inode = s.inode, .size = s.size, .mtime = s.mtime.toNanoseconds(), .ctime = s.ctime.toNanoseconds() };
+        return .{ .device = 0, .inode = @as(u64, @bitCast(s.inode)), .size = s.size, .mtime = s.mtime.toNanoseconds(), .ctime = s.ctime.toNanoseconds() };
     }
 
     pub fn equal(a: Key, b: Key) bool {
diff --git a/zig/src/core/picture/picture_test.zig b/zig/src/core/picture/picture_test.zig
index 287493a71c..80ddb5f4d1 100644
--- a/zig/src/core/picture/picture_test.zig
+++ b/zig/src/core/picture/picture_test.zig
@@ -119,7 +119,7 @@ test "edge pictures (zig/tests/fixtures/picture/edges.json) get 0.6.6's outcome,
 /// WebP cases need the pinned libwebp: required when TENSORFOLD_LIBWEBP names one, skipped when none loads.
 fn webpReady(a: std.mem.Allocator) !bool {
     if (picture.webp.available(std.testing.io, a)) return true;
-    if (std.testing.environ.getPosix(picture.webp.env) != null) return error.PinnedLibraryDidNotLoad;
+    if (@import("sys").getenv(picture.webp.env) != null) return error.PinnedLibraryDidNotLoad;
     std.debug.print("libwebp {s} not found: WebP pixel cases skipped\n", .{picture.webp.pinned});
     return false;
 }
diff --git a/zig/src/core/picture/webp.zig b/zig/src/core/picture/webp.zig
index db6d9f7131..b45a2180e1 100644
--- a/zig/src/core/picture/webp.zig
+++ b/zig/src/core/picture/webp.zig
@@ -66,7 +66,7 @@ fn find(io: std.Io, a: Allocator) Loaded {
 
 /// One candidate: null when it does not load or lacks a function; `.wrong` when its version is not the pin.
 fn open(path: []const u8) ?Loaded {
-    var dl = std.DynLib.open(path) catch return null;
+    var dl = @import("sys").DynLib.open(path) catch return null;
     var f: Fns = undefined;
     const info = @typeInfo(Fns).@"struct";
     inline for (info.field_names, info.field_types) |name, T| {
diff --git a/zig/src/cuda/copy_source_close_test.zig b/zig/src/cuda/copy_source_close_test.zig
index d1bf218d06..f71959f230 100644
--- a/zig/src/cuda/copy_source_close_test.zig
+++ b/zig/src/cuda/copy_source_close_test.zig
@@ -84,7 +84,7 @@ pub fn run(reap: *const fn (*const Driver) void) !void {
     Fake.s = .{};
     Fake.reap = reap;
     var d: Driver = undefined;
-    d.lib = try std.DynLib.open(if (builtin.os.tag == .windows) "kernel32.dll" else if (builtin.os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
+    d.lib = try @import("sys").DynLib.open(if (builtin.os.tag == .windows) "kernel32.dll" else if (builtin.os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
     d.api.cuMemHostAlloc = Fake.alloc;
     d.api.cuMemFreeHost = Fake.free;
     d.api.cuMemcpyHtoDAsync_v2 = Fake.copy;
diff --git a/zig/src/cuda/copy_source_race_test.zig b/zig/src/cuda/copy_source_race_test.zig
index 0eed0c9a1b..68583dfb15 100644
--- a/zig/src/cuda/copy_source_race_test.zig
+++ b/zig/src/cuda/copy_source_race_test.zig
@@ -97,7 +97,7 @@ pub fn run(reaper: *const fn (*const Driver) void, completing: bool) !void {
     Fake.reaper = reaper;
     Fake.completing = completing;
     var d: Driver = undefined;
-    d.lib = try std.DynLib.open(if (builtin.os.tag == .windows) "kernel32.dll" else if (builtin.os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
+    d.lib = try @import("sys").DynLib.open(if (builtin.os.tag == .windows) "kernel32.dll" else if (builtin.os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
     d.api.cuMemHostAlloc = Fake.alloc;
     d.api.cuMemFreeHost = Fake.free;
     d.api.cuMemcpyHtoDAsync_v2 = Fake.copy;
diff --git a/zig/src/cuda/memory_receipt.zig b/zig/src/cuda/memory_receipt.zig
index a9b6aa7060..6143868f9e 100644
--- a/zig/src/cuda/memory_receipt.zig
+++ b/zig/src/cuda/memory_receipt.zig
@@ -1,13 +1,14 @@
 //! Opt-in process-local allocator boundary receipts; no CUDA calls and no global peak reset unless requested.
 const std = @import("std");
+const platform = @import("sys");
 const memory = @import("memory.zig");
 
 pub const Phase = enum { load_begin, load_end, cleanup_begin, cleanup_end, reset };
 
 pub fn enabled() bool {
     if (@import("builtin").is_test) return false;
-    const env = std.c.getenv("TF_CUDA_MEMORY_RECEIPTS") orelse return false;
-    return std.mem.eql(u8, std.mem.span(env), "1");
+    const env = platform.getenv("TF_CUDA_MEMORY_RECEIPTS") orelse return false;
+    return std.mem.eql(u8, env, "1");
 }
 
 pub fn emit(rank: u32, phase: Phase, reset: bool) void {
@@ -17,22 +18,22 @@ pub fn emit(rank: u32, phase: Phase, reset: bool) void {
 
 pub fn emitUsage(rank: u32, phase: Phase, reset: bool, usage: memory.Usage) void {
     if (!enabled()) return;
-    var ts: std.c.timespec = undefined;
-    const ns: ?u64 = if (std.c.clock_gettime(.MONOTONIC, &ts) == 0)
+    var ts: platform.c.timespec = undefined;
+    const ns: ?u64 = if (platform.c.clock_gettime(platform.c.CLOCK.MONOTONIC, &ts) == 0)
         @as(u64, @intCast(ts.sec)) * std.time.ns_per_s + @as(u64, @intCast(ts.nsec))
     else
         null;
-    const receipt: Receipt = .{ .pid = std.c.getpid(), .rank = rank, .phase = phase, .reset = reset, .monotonic_ns = ns, .usage = usage };
+    const receipt: Receipt = .{ .pid = platform.c.getpid(), .rank = rank, .phase = phase, .reset = reset, .monotonic_ns = ns, .usage = usage };
     std.debug.print("TF_MEMORY {f}\n", .{std.json.fmt(receipt, .{})});
 }
 
 pub fn snapshot(rank: u32, phase: Phase, reset: bool, monotonic_ns: ?u64) Receipt {
-    return .{ .pid = std.c.getpid(), .rank = rank, .phase = phase, .reset = reset, .monotonic_ns = monotonic_ns, .usage = memory.usage(reset) };
+    return .{ .pid = platform.c.getpid(), .rank = rank, .phase = phase, .reset = reset, .monotonic_ns = monotonic_ns, .usage = memory.usage(reset) };
 }
 
 pub const Receipt = struct {
     schema: u32 = 1,
-    pid: std.c.pid_t,
+    pid: platform.c.pid_t,
     rank: u32,
     phase: Phase,
     reset: bool,
diff --git a/zig/src/cuda/roce.zig b/zig/src/cuda/roce.zig
index 25295d9bcc..9faf1af51a 100644
--- a/zig/src/cuda/roce.zig
+++ b/zig/src/cuda/roce.zig
@@ -1,6 +1,7 @@
 //! Two-rank all-gathers as one-shot RoCE writes after b12x's RoCEnante (local-inference-lab/b12x, Apache-2.0).
 
 const std = @import("std");
+const platform = @import("sys");
 const abi = @import("abi.zig");
 const Driver = @import("driver.zig").Driver;
 const memory = @import("memory.zig");
@@ -212,7 +213,7 @@ const Hca = struct {
 pub const Gather = struct {
     gpa: std.mem.Allocator,
     d: *const Driver,
-    lib: std.DynLib,
+    lib: platform.DynLib,
     api: Api,
     rank: u8,
     hcas: [max_hcas]Hca = @splat(.{}),
@@ -284,7 +285,7 @@ pub const Gather = struct {
 
     fn setup(g: *Gather, gpa: std.mem.Allocator, io: std.Io, d: *const Driver, rank: u8, max_bytes: usize) !void {
         g.* = .{ .gpa = gpa, .d = d, .lib = undefined, .api = undefined, .rank = rank, .layout = .of(max_bytes), .region = undefined, .region_dev = 0, .counters = undefined, .module = undefined, .f = undefined, .max_bytes = max_bytes };
-        g.lib = std.DynLib.open("libibverbs.so.1") catch return error.VerbsUnavailable;
+        g.lib = platform.DynLib.open("libibverbs.so.1") catch return error.VerbsUnavailable;
         errdefer g.lib.close();
         const info = @typeInfo(Api).@"struct";
         inline for (info.field_names, info.field_types) |name, T| {
diff --git a/zig/src/families/drama3/cuda/blas.zig b/zig/src/families/drama3/cuda/blas.zig
index 60ff29a991..97a78c72c5 100644
--- a/zig/src/families/drama3/cuda/blas.zig
+++ b/zig/src/families/drama3/cuda/blas.zig
@@ -1,5 +1,6 @@
 //! cuBLAS (the classic API) opened at run time: the strided batched GEMM Drama 3's prompt rows and codec take.
 const std = @import("std");
+const platform = @import("sys");
 const cuda = @import("cuda");
 
 pub const Status = c_int;
@@ -33,7 +34,7 @@ pub const Api = struct {
 
 /// The library and one handle with its workspace, bound to a stream per call.
 pub const Blas = struct {
-    lib: std.DynLib,
+    lib: platform.DynLib,
     api: Api,
     h: Handle,
 
@@ -49,7 +50,7 @@ pub const Blas = struct {
     }
 
     fn openPath(path: []const u8, workspace: Ptr, workspace_bytes: usize) Error!Blas {
-        var lib = std.DynLib.open(path) catch return error.LibraryUnavailable;
+        var lib = platform.DynLib.open(path) catch return error.LibraryUnavailable;
         errdefer lib.close();
         var b: Blas = .{ .lib = lib, .api = undefined, .h = null };
         const info = @typeInfo(Api).@"struct";
diff --git a/zig/src/families/flashnext/cuda_lanes_restore_test.zig b/zig/src/families/flashnext/cuda_lanes_restore_test.zig
index e0e5967ecf..aa441f6166 100644
--- a/zig/src/families/flashnext/cuda_lanes_restore_test.zig
+++ b/zig/src/families/flashnext/cuda_lanes_restore_test.zig
@@ -49,6 +49,7 @@ pub fn fakeDriver() cuda.Driver {
 
 /// Exercise the actual lane open and explicit snapshot entry points with a CPU control peer.
 pub fn run(comptime Cuda: type, comptime open: anytype, comptime restore: anytype, comptime drop: anytype, explicit: bool, drafting: bool) !void {
+    if (@import("builtin").os.tag == .windows) return error.SkipZigTest; // the peer rank is a Unix socket pair
     const a = std.testing.allocator;
     Device.next = 0x100000;
     Device.copies = 0;
diff --git a/zig/src/families/flashnext/cuda_ple.zig b/zig/src/families/flashnext/cuda_ple.zig
index 459f56905d..1ea1fe4e7b 100644
--- a/zig/src/families/flashnext/cuda_ple.zig
+++ b/zig/src/families/flashnext/cuda_ple.zig
@@ -388,7 +388,7 @@ pub const Ngram = struct {
             const file = findOpen(g.files.items, path) orelse file: {
                 var o = try Open.load(gpa, io, path);
                 errdefer o.close(gpa);
-                std.posix.madvise(o.file.map.memory.ptr, o.file.map.memory.len, std.posix.MADV.RANDOM) catch {};
+                if (@import("builtin").os.tag != .windows) std.posix.madvise(o.file.map.memory.ptr, o.file.map.memory.len, std.posix.MADV.RANDOM) catch {};
                 try g.files.append(gpa, o);
                 break :file &g.files.items[g.files.items.len - 1];
             };
diff --git a/zig/src/families/flashnext/cuda_rank_media.zig b/zig/src/families/flashnext/cuda_rank_media.zig
index 79a783b8b4..4461af16e0 100644
--- a/zig/src/families/flashnext/cuda_rank_media.zig
+++ b/zig/src/families/flashnext/cuda_rank_media.zig
@@ -58,5 +58,5 @@ test "rank media frames are bounded and rebuild the same image plan" {
 }
 
 test {
-    _ = @import("cuda_rank_media_test.zig");
+    if (@import("builtin").os.tag != .windows) _ = @import("cuda_rank_media_test.zig"); // over a Unix socket pair
 }
diff --git a/zig/src/families/flashnext/cuda_rank_wire.zig b/zig/src/families/flashnext/cuda_rank_wire.zig
index a49429688d..98606b2994 100644
--- a/zig/src/families/flashnext/cuda_rank_wire.zig
+++ b/zig/src/families/flashnext/cuda_rank_wire.zig
@@ -1,6 +1,7 @@
 //! Bounded Linux TCP control channel. Tensor bytes travel over NCCL; this channel owns command order and peer EOF.
 const std = @import("std");
 const builtin = @import("builtin");
+const platform = @import("sys");
 const p = std.posix;
 const net = std.Io.net;
 
@@ -43,8 +44,8 @@ pub const Header = struct {
 };
 
 pub fn now() u64 {
-    var ts: std.c.timespec = undefined;
-    if (std.c.clock_gettime(.MONOTONIC, &ts) != 0 or ts.sec < 0) return 0;
+    var ts: platform.c.timespec = undefined;
+    if (platform.c.clock_gettime(platform.c.CLOCK.MONOTONIC, &ts) != 0 or ts.sec < 0) return 0;
     return @as(u64, @intCast(ts.sec)) * std.time.ns_per_ms * 1000 + @as(u64, @intCast(ts.nsec));
 }
 
@@ -53,8 +54,8 @@ pub fn deadline(ms: u64) u64 {
 }
 
 pub fn pause() void {
-    const ts: std.c.timespec = .{ .sec = 0, .nsec = 50_000 };
-    _ = std.c.nanosleep(&ts, null);
+    const ts: platform.c.timespec = .{ .sec = 0, .nsec = 50_000 };
+    _ = platform.c.nanosleep(&ts, null);
 }
 
 fn wait(fd: p.socket_t, events: i16, until: ?u64) !void {
@@ -131,6 +132,7 @@ pub const Link = struct {
     }
 
     pub fn send(l: Link, bytes: []const u8, until: u64) !void {
+        if (builtin.os.tag == .windows) return error.RankPlatform; // no MSG.NOSIGNAL or DONTWAIT in Winsock
         var at: usize = 0;
         while (at < bytes.len) {
             if (now() >= until) return error.RankTimeout;
@@ -148,6 +150,7 @@ pub const Link = struct {
     }
 
     pub fn receive(l: Link, bytes: []u8, until: ?u64) !void {
+        if (builtin.os.tag == .windows) return error.RankPlatform;
         var at: usize = 0;
         while (at < bytes.len) {
             if (until) |end| if (now() >= end) return error.RankTimeout;
@@ -179,6 +182,7 @@ pub const Link = struct {
 
     /// A non-consuming EOF check while a GPU collective progresses. A future control frame stays queued.
     pub fn checkPeer(l: Link) !void {
+        if (builtin.os.tag == .windows) return error.RankPlatform;
         var byte: u8 = 0;
         const rc = std.c.recv(l.fd, &byte, 1, p.MSG.PEEK | p.MSG.DONTWAIT);
         if (rc == 0) return error.RankDisconnected;
diff --git a/zig/src/families/nemotron/cuda_pages_safety_test.zig b/zig/src/families/nemotron/cuda_pages_safety_test.zig
index 057c443672..3a00438422 100644
--- a/zig/src/families/nemotron/cuda_pages_safety_test.zig
+++ b/zig/src/families/nemotron/cuda_pages_safety_test.zig
@@ -159,6 +159,7 @@ const Child = struct {
     extern "c" fn fork() c_int;
 };
 test "kv safety: paged constructor fatal completion failure precedes every destination or parent release" {
+    if (@import("builtin").os.tag == .windows) return error.SkipZigTest; // the child process comes from fork
     for ([_]Case{ .clear, .slots, .cpu, .wait }) |case| {
         const pid = Child.fork();
         if (pid < 0) return error.ForkFailed;
diff --git a/zig/src/native/cuda_load_gpu.zig b/zig/src/native/cuda_load_gpu.zig
index 41d8abd90e..1774c6eebb 100644
--- a/zig/src/native/cuda_load_gpu.zig
+++ b/zig/src/native/cuda_load_gpu.zig
@@ -156,7 +156,7 @@ const Fake = struct {
         const g = try t.allocator.create(Gpu);
         g.gpa = t.allocator;
         g.next = null;
-        g.driver.lib = try std.DynLib.open(if (@import("builtin").os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
+        g.driver.lib = try @import("sys").DynLib.open(if (@import("builtin").os.tag == .windows) "kernel32.dll" else if (@import("builtin").os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
         g.driver.api.cuMemAlloc_v2 = alloc;
         g.driver.api.cuMemFree_v2 = free;
         g.driver.api.cuMemcpyHtoD_v2 = copy;
@@ -278,7 +278,7 @@ const Barrier = @import("load_barrier_test.zig");
 fn secondGpu(first: *Gpu) !*Gpu {
     const g = try t.allocator.create(Gpu);
     g.* = first.*;
-    g.driver.lib = try std.DynLib.open(if (@import("builtin").os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
+    g.driver.lib = try @import("sys").DynLib.open(if (@import("builtin").os.tag == .windows) "kernel32.dll" else if (@import("builtin").os.tag == .macos) "/usr/lib/libSystem.B.dylib" else "libc.so.6");
     g.ctx = .{ .d = &g.driver, .device = 1, .handle = @ptrFromInt(7) };
     g.next = null;
     return g;
diff --git a/zig/src/server/listener.zig b/zig/src/server/listener.zig
index 55dee3333c..4095cd8e51 100644
--- a/zig/src/server/listener.zig
+++ b/zig/src/server/listener.zig
@@ -155,10 +155,12 @@ test "accepted IPv4 and IPv6 TCP streams disable Nagle for incremental responses
         if (posix.errno(posix.system.connect(client, &storage.any, len)) != .SUCCESS) return error.ConnectFailed;
         const accepted = acceptWithin(listener, 2000) orelse return error.AcceptFailed;
         defer _ = posix.system.close(accepted.fd);
-        var value: c_int = 0;
-        var size: posix.socklen_t = @sizeOf(c_int);
-        try std.testing.expect(std.c.getsockopt(accepted.fd, posix.IPPROTO.TCP, posix.TCP.NODELAY, &value, &size) == 0);
-        try std.testing.expect(value != 0);
+        if (!platform.windows) { // sys wraps no getsockopt to read the option back on Windows
+            var value: c_int = 0;
+            var size: posix.socklen_t = @sizeOf(c_int);
+            try std.testing.expect(std.c.getsockopt(accepted.fd, posix.IPPROTO.TCP, posix.TCP.NODELAY, &value, &size) == 0);
+            try std.testing.expect(value != 0);
+        }
     }
 }
 
diff --git a/zig/src/server/root.zig b/zig/src/server/root.zig
index c7b31b5004..dcb194bf17 100644
--- a/zig/src/server/root.zig
+++ b/zig/src/server/root.zig
@@ -39,7 +39,7 @@ test {
     _ = @import("chunk_plan.zig");
     _ = @import("status_routes.zig");
     _ = @import("speech.zig");
-    _ = @import("speech_stream_test.zig");
+    if (comptime @import("builtin").os.tag != .windows) _ = @import("speech_stream_test.zig"); // over a Unix socket pair
     _ = @import("drama_prompt.zig");
     _ = @import("health_test.zig");
     _ = @import("memory_routes.zig");
@@ -110,9 +110,6 @@ test {
         _ = @import("benchmark_cli_test.zig");
         _ = @import("benchmark_run_test.zig");
         _ = @import("benchmark_early_test.zig");
+        _ = @import("benchmark_config_test.zig");
     }
 }
-
-test {
-    _ = @import("benchmark_config_test.zig");
-}
diff --git a/zig/src/server/speech.zig b/zig/src/server/speech.zig
index ec6166bed9..b7f474facd 100644
--- a/zig/src/server/speech.zig
+++ b/zig/src/server/speech.zig
@@ -1,5 +1,6 @@
 //! POST /v1/audio/speech: WAV or raw PCM whole, or raw PCM chunks with stream; only complete synthesis is audio.
 const std = @import("std");
+const platform = @import("sys");
 const api = @import("engine_api");
 const json = @import("json");
 const errors = @import("errors.zig");
@@ -276,10 +277,10 @@ fn startStream(c: *Conn, sample_rate: u32) error{Closed}!void {
     if (c.close) c.addHeader("Connection", "close") catch return error.Closed;
     // each chunk leaves at once rather than waiting on the client's acknowledgement of the last (no effect off TCP)
     const one: c_int = 1;
-    _ = std.c.setsockopt(c.fd, std.posix.IPPROTO.TCP, std.posix.TCP.NODELAY, &one, @sizeOf(c_int));
+    if (platform.windows) platform.posix.setsockopt(c.fd, platform.posix.IPPROTO.TCP, platform.posix.TCP.NODELAY, std.mem.asBytes(&one)) catch {} else _ = std.c.setsockopt(c.fd, std.posix.IPPROTO.TCP, std.posix.TCP.NODELAY, &one, @sizeOf(c_int));
     // a blocking write into a full socket returns after the stall limit, so writeAll's poll can end a stalled stream
-    const stall: std.posix.timeval = .{ .sec = @divTrunc(c.timeouts.write_ms, 1000), .usec = @rem(c.timeouts.write_ms, 1000) * 1000 };
-    std.posix.setsockopt(c.fd, std.posix.SOL.SOCKET, std.posix.SO.SNDTIMEO, std.mem.asBytes(&stall)) catch {};
+    const stall: platform.posix.timeval = .{ .sec = @divTrunc(c.timeouts.write_ms, 1000), .usec = @rem(c.timeouts.write_ms, 1000) * 1000 };
+    platform.posix.setsockopt(c.fd, platform.posix.SOL.SOCKET, platform.posix.SO.SNDTIMEO, std.mem.asBytes(&stall)) catch {};
     return c.finish("");
 }
 
diff --git a/zig/src/vision/qwen_cuda_test.zig b/zig/src/vision/qwen_cuda_test.zig
index 113cee6323..a088bb4069 100644
--- a/zig/src/vision/qwen_cuda_test.zig
+++ b/zig/src/vision/qwen_cuda_test.zig
@@ -245,6 +245,7 @@ const Child = struct {
     extern "c" fn kill(c_int, c_int) c_int;
 };
 test "a stream and context that never complete end the process before an owner is freed" {
+    if (@import("builtin").os.tag == .windows) return error.SkipZigTest; // the child process comes from fork
     const pid = Child.fork();
     if (pid < 0) return error.ForkFailed;
     if (pid == 0) {