diff --git a/internal/cache/s3.go b/internal/cache/s3.go
index f2a029f..1206456 100644
--- a/internal/cache/s3.go
+++ b/internal/cache/s3.go
@@ -262,13 +262,13 @@ func (s *S3) Open(ctx context.Context, key Key) (io.ReadCloser, http.Header, err
 		}
 	}
 
-	// Get object
-	obj, err := s.client.GetObject(ctx, s.config.Bucket, objectName, minio.GetObjectOptions{})
+	// Download object using parallel range-GET for large objects.
+	reader, err := s.parallelGetReader(ctx, s.config.Bucket, objectName, objInfo.Size)
 	if err != nil {
-		return nil, nil, errors.Errorf("failed to get object: %w", err)
+		return nil, nil, err
 	}
 
-	return &s3Reader{obj: obj}, headers, nil
+	return reader, headers, nil
 }
 
 // refreshExpiration updates the Expires-At metadata on an S3 object using
diff --git a/internal/cache/s3_parallel_get.go b/internal/cache/s3_parallel_get.go
new file mode 100644
index 0000000..b12dbb9
--- /dev/null
+++ b/internal/cache/s3_parallel_get.go
@@ -0,0 +1,116 @@
+package cache
+
+import (
+	"context"
+	"io"
+	"sync"
+
+	"github.com/alecthomas/errors"
+	"github.com/minio/minio-go/v7"
+)
+
+const (
+	// s3DownloadChunkSize is the size of each parallel range-GET request.
+	// 32 MiB matches the gradle-cache-tool's benchmarked default.
+	s3DownloadChunkSize = 32 << 20
+	// s3DownloadWorkers is the number of concurrent range-GET requests.
+	// Benchmarking showed no throughput difference from 4 to 128 workers
+	// (extraction IOPS is the bottleneck), so 8 keeps connection count low.
+	s3DownloadWorkers = 8
+)
+
+// parallelGetReader returns an io.ReadCloser that downloads the S3 object
+// using parallel range-GET requests and reassembles chunks in order.
+// For objects smaller than one chunk, it falls back to a single GetObject.
+func (s *S3) parallelGetReader(ctx context.Context, bucket, objectName string, size int64) (io.ReadCloser, error) {
+	if size <= s3DownloadChunkSize {
+		// Small object: single stream.
+		obj, err := s.client.GetObject(ctx, bucket, objectName, minio.GetObjectOptions{})
+		if err != nil {
+			return nil, errors.Errorf("failed to get object: %w", err)
+		}
+		return &s3Reader{obj: obj}, nil
+	}
+
+	// Large object: parallel range requests reassembled in order via io.Pipe.
+	pr, pw := io.Pipe()
+	go func() {
+		pw.CloseWithError(s.parallelGet(ctx, bucket, objectName, size, pw))
+	}()
+	return pr, nil
+}
+
+// parallelGet downloads an S3 object in parallel chunks and writes them in
+// order to w. Each worker downloads its chunk into memory so the TCP
+// connection stays active at full speed. Peak memory: numWorkers × chunkSize.
+func (s *S3) parallelGet(ctx context.Context, bucket, objectName string, size int64, w io.Writer) error {
+	numChunks := int((size + s3DownloadChunkSize - 1) / s3DownloadChunkSize)
+	numWorkers := min(s3DownloadWorkers, numChunks)
+
+	type chunkResult struct {
+		data []byte
+		err  error
+	}
+
+	// One buffered channel per chunk so workers never block after reading.
+	results := make([]chan chunkResult, numChunks)
+	for i := range results {
+		results[i] = make(chan chunkResult, 1)
+	}
+
+	// Work queue of chunk indices.
+	work := make(chan int, numChunks)
+	for i := range numChunks {
+		work <- i
+	}
+	close(work)
+
+	var wg sync.WaitGroup
+	for range numWorkers {
+		wg.Go(func() {
+			for seq := range work {
+				start := int64(seq) * s3DownloadChunkSize
+				end := min(start+s3DownloadChunkSize-1, size-1)
+
+				opts := minio.GetObjectOptions{}
+				if err := opts.SetRange(start, end); err != nil {
+					results[seq] <- chunkResult{err: errors.Errorf("set range %d-%d: %w", start, end, err)}
+					continue
+				}
+
+				obj, err := s.client.GetObject(ctx, bucket, objectName, opts)
+				if err != nil {
+					results[seq] <- chunkResult{err: errors.Errorf("get range %d-%d: %w", start, end, err)}
+					continue
+				}
+
+				// Drain the body immediately so the TCP connection stays at
+				// full speed. All workers do this concurrently, saturating
+				// the available S3 bandwidth.
+				data, readErr := io.ReadAll(obj)
+				obj.Close() //nolint:errcheck,gosec
+				results[seq] <- chunkResult{data: data, err: readErr}
+			}
+		})
+	}
+
+	// Write chunks in order. Each receive blocks until that chunk's worker
+	// finishes, while other workers continue downloading concurrently.
+	var writeErr error
+	for _, ch := range results {
+		r := <-ch
+		if writeErr != nil {
+			continue // drain remaining channels so goroutines can exit
+		}
+		if r.err != nil {
+			writeErr = r.err
+			continue
+		}
+		if _, err := w.Write(r.data); err != nil {
+			writeErr = err
+		}
+	}
+
+	wg.Wait()
+	return writeErr
+}
diff --git a/internal/snapshot/snapshot.go b/internal/snapshot/snapshot.go
index a3147e8..ee9c4e8 100644
--- a/internal/snapshot/snapshot.go
+++ b/internal/snapshot/snapshot.go
@@ -2,6 +2,7 @@
 package snapshot
 
 import (
+	"archive/tar"
 	"bufio"
 	"bytes"
 	"context"
@@ -12,6 +13,8 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"strings"
+	"sync"
 	"time"
 
 	"github.com/alecthomas/errors"
@@ -187,9 +190,32 @@ func Restore(ctx context.Context, remote cache.Cache, key cache.Key, directory s
 	return Extract(ctx, rc, directory, threads)
 }
 
+const (
+	// extractWorkers is the number of goroutines writing files concurrently
+	// during parallel tar extraction. Hides per-file open/write/close syscall
+	// latency so the tar-stream reader (and download pipeline behind it) is not
+	// stalled waiting for individual file writes to complete.
+	// Benchmarked on r8id.metal-48xlarge (NVMe, 96 cores) with a 334K-file
+	// bundle: 64 workers = 6.27s, 128 = 6.84s (extra GC pressure outweighs
+	// any I/O concurrency gain).
+	extractWorkers = 64
+	// maxParallelFileSize is the largest file that will be buffered in memory
+	// and dispatched to the worker pool. Files larger than this are written
+	// inline in the main goroutine to keep peak memory bounded.
+	// At 4 MiB, 99.97% of Gradle cache entries go through the parallel path.
+	maxParallelFileSize = 4 << 20 // 4 MiB
+)
+
 // Extract decompresses a zstd+tar stream into directory, preserving all file
 // permissions, ownership, and symlinks. threads controls zstd parallelism;
 // 0 uses all available CPU cores.
+//
+// The single-threaded bottleneck on restore is writing files to disk. Even
+// though tar entries must be read sequentially (the format has no index), the
+// actual file writes are independent. The extractor dispatches each entry
+// (buffered in memory, ≤4 MiB) to one of 64 worker goroutines that write
+// concurrently. This hides the per-file syscall latency (~20µs × N files)
+// behind parallelism.
 func Extract(ctx context.Context, r io.Reader, directory string, threads int) error {
 	if threads <= 0 {
 		threads = runtime.NumCPU()
@@ -212,15 +238,193 @@ func Extract(ctx context.Context, r io.Reader, directory string, threads int) er
 	}
 	defer dec.Close()
 
-	tarCmd := exec.CommandContext(ctx, "tar", "-xpf", "-", "-C", directory)
-	tarCmd.Stdin = dec
+	return extractTarParallel(ctx, dec, directory)
+}
 
-	var tarStderr bytes.Buffer
-	tarCmd.Stderr = &tarStderr
+type writeJob struct {
+	target string
+	mode   os.FileMode
+	data   []byte
+}
+
+// safePath validates that name is a relative path that stays within dir when
+// joined. It rejects absolute paths and parent traversals (".."). Returns the
+// resolved path under dir.
+func safePath(dir, name string) (string, error) {
+	clean := filepath.Clean(name)
+	if filepath.IsAbs(clean) {
+		return "", errors.Errorf("path %q is absolute", name)
+	}
+	if clean == ".." || strings.HasPrefix(clean, ".."+string(os.PathSeparator)) {
+		return "", errors.Errorf("path %q escapes destination directory", name)
+	}
+	joined := filepath.Join(dir, clean)
+	if !strings.HasPrefix(joined, dir+string(os.PathSeparator)) && joined != dir {
+		return "", errors.Errorf("path %q resolves outside destination directory", name)
+	}
+	return joined, nil
+}
+
+// extractTarParallel reads a tar stream and writes files using a pool of
+// goroutines. The main goroutine reads tar entries and buffers small file
+// contents; workers write those files to disk concurrently. Large files are
+// written inline to keep memory use bounded.
+func extractTarParallel(ctx context.Context, r io.Reader, dir string) error {
+	// Resolve dir to absolute so containment checks are reliable.
+	var err error
+	dir, err = filepath.Abs(dir)
+	if err != nil {
+		return errors.Wrap(err, "resolve destination directory")
+	}
+
+	jobs := make(chan writeJob, extractWorkers*2)
+
+	var (
+		wg           sync.WaitGroup
+		writeErrOnce sync.Once
+		writeErr     error
+	)
+
+	for range extractWorkers {
+		wg.Go(func() {
+			for job := range jobs {
+				f, err := os.OpenFile(job.target, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, job.mode)
+				if err != nil {
+					writeErrOnce.Do(func() { writeErr = errors.Errorf("open %s: %w", filepath.Base(job.target), err) })
+					continue
+				}
+				if _, err := f.Write(job.data); err != nil {
+					f.Close() //nolint:errcheck,gosec
+					writeErrOnce.Do(func() { writeErr = errors.Errorf("write %s: %w", filepath.Base(job.target), err) })
+					continue
+				}
+				if err := f.Close(); err != nil {
+					writeErrOnce.Do(func() { writeErr = errors.Errorf("close %s: %w", filepath.Base(job.target), err) })
+				}
+			}
+		})
+	}
+
+	copyBuf := make([]byte, 1<<20) // reused only for inline large-file writes
+
+	// createdDirs is accessed only by the main goroutine, so no mutex needed.
+	createdDirs := make(map[string]struct{})
+	ensureDir := func(d string, mode os.FileMode) error {
+		if _, ok := createdDirs[d]; ok {
+			return nil
+		}
+		if err := os.MkdirAll(d, mode); err != nil { //nolint:gosec // path is validated by caller
+			return errors.Wrap(err, "mkdir")
+		}
+		createdDirs[d] = struct{}{}
+		return nil
+	}
 
-	if err := tarCmd.Run(); err != nil {
-		return errors.Errorf("tar failed: %w: %s", err, tarStderr.String())
+	tr := tar.NewReader(r)
+	var readErr error
+loop:
+	for {
+		if err := ctx.Err(); err != nil {
+			readErr = errors.Wrap(err, "context cancelled")
+			break
+		}
+
+		hdr, err := tr.Next()
+		if errors.Is(err, io.EOF) {
+			break
+		}
+		if err != nil {
+			readErr = errors.Wrap(err, "read tar entry")
+			break
+		}
+
+		target, err := safePath(dir, hdr.Name)
+		if err != nil {
+			readErr = errors.Errorf("unsafe tar entry %q: %w", hdr.Name, err)
+			break
+		}
+
+		switch hdr.Typeflag {
+		case tar.TypeDir:
+			if err := ensureDir(target, hdr.FileInfo().Mode()); err != nil {
+				readErr = errors.Errorf("mkdir %s: %w", hdr.Name, err)
+				break loop
+			}
+
+		case tar.TypeLink:
+			if err := ensureDir(filepath.Dir(target), 0o755); err != nil {
+				readErr = errors.Errorf("mkdir for hardlink %s: %w", hdr.Name, err)
+				break loop
+			}
+			linkTarget, err := safePath(dir, hdr.Linkname)
+			if err != nil {
+				readErr = errors.Errorf("unsafe hardlink target %q: %w", hdr.Linkname, err)
+				break loop
+			}
+			if err := os.Link(linkTarget, target); err != nil {
+				readErr = errors.Errorf("hardlink %s → %s: %w", hdr.Name, hdr.Linkname, err)
+				break loop
+			}
+
+		case tar.TypeReg:
+			if err := ensureDir(filepath.Dir(target), 0o755); err != nil {
+				readErr = errors.Errorf("mkdir %s: %w", hdr.Name, err)
+				break loop
+			}
+
+			if hdr.Size <= maxParallelFileSize {
+				// Buffer in memory and dispatch to worker pool so the main
+				// goroutine can continue reading the tar stream immediately.
+				buf := make([]byte, hdr.Size)
+				if _, err := io.ReadFull(tr, buf); err != nil {
+					readErr = errors.Errorf("read %s: %w", hdr.Name, err)
+					break loop
+				}
+				// Propagate worker errors early.
+				if writeErr != nil {
+					readErr = writeErr
+					break loop
+				}
+				jobs <- writeJob{target: target, mode: hdr.FileInfo().Mode(), data: buf}
+			} else {
+				// Large file: write inline to keep memory bounded.
+				f, err := os.OpenFile(target, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, hdr.FileInfo().Mode()) //nolint:gosec // path traversal guarded above
+				if err != nil {
+					readErr = errors.Errorf("open %s: %w", hdr.Name, err)
+					break loop
+				}
+				if _, err := io.CopyBuffer(f, io.LimitReader(tr, hdr.Size), copyBuf); err != nil {
+					f.Close() //nolint:errcheck,gosec
+					readErr = errors.Errorf("write %s: %w", hdr.Name, err)
+					break loop
+				}
+				if err := f.Close(); err != nil {
+					readErr = errors.Errorf("close %s: %w", hdr.Name, err)
+					break loop
+				}
+			}
+
+		case tar.TypeSymlink:
+			if err := ensureDir(filepath.Dir(target), 0o755); err != nil {
+				readErr = errors.Errorf("mkdir for symlink %s: %w", hdr.Name, err)
+				break loop
+			}
+			if _, err := safePath(dir, hdr.Linkname); err != nil {
+				readErr = errors.Errorf("unsafe symlink target %q: %w", hdr.Linkname, err)
+				break loop
+			}
+			if err := os.Symlink(hdr.Linkname, target); err != nil {
+				readErr = errors.Errorf("symlink %s → %s: %w", hdr.Name, hdr.Linkname, err)
+				break loop
+			}
+		}
 	}
 
-	return nil
+	close(jobs)
+	wg.Wait()
+
+	if readErr != nil {
+		return readErr
+	}
+	return writeErr
 }