diff --git a/src/models.rs b/src/models.rs
index 4144c71..8139f19 100644
--- a/src/models.rs
+++ b/src/models.rs
@@ -14,7 +14,7 @@ use reqwest::blocking::Client;
 use reqwest::redirect::Policy;
 use serde::Deserialize;
 use sha2::{Digest, Sha256};
-use indicatif::{ProgressBar, ProgressStyle};
+use indicatif::{ProgressBar, ProgressStyle, MultiProgress};
 use atty::Stream;
 
 // --- Model downloader: list & download ggml models from Hugging Face ---
@@ -496,16 +496,55 @@ pub fn run_interactive_model_downloader() -> Result<()> {
         qlog!("No selection. Aborting download.");
         return Ok(());
     }
-    for m in selected {
-        if let Err(e) = download_one_model(&client, models_dir, &m) {
-            elog!("Error: {:#}", e);
+
+    // Parallel downloads with bounded concurrency. Default 3; override via POLYSCRIBE_MAX_PARALLEL_DOWNLOADS (1..=6).
+    let max_jobs = std::env::var("POLYSCRIBE_MAX_PARALLEL_DOWNLOADS")
+        .ok()
+        .and_then(|s| s.parse::<usize>().ok())
+        .map(|n| n.clamp(1, 6))
+        .unwrap_or(3);
+
+    // Use a MultiProgress to render per-model bars concurrently when interactive.
+    let mp_opt = if !crate::is_quiet() && !crate::is_no_progress() && atty::is(Stream::Stderr) {
+        Some(MultiProgress::new())
+    } else {
+        None
+    };
+
+    let mut i = 0;
+    while i < selected.len() {
+        let end = std::cmp::min(i + max_jobs, selected.len());
+        let mut handles = Vec::new();
+        for m in selected[i..end].iter().cloned() {
+            let client2 = client.clone();
+            let models_dir2 = models_dir.to_path_buf();
+            let pb_opt = if let Some(mp) = &mp_opt {
+                let pb = mp.add(ProgressBar::new(m.size));
+                let style = ProgressStyle::with_template("Downloading {prefix} ({total_bytes}) [{bar:40.cyan/blue}] {bytes}/{total_bytes} ({percent}%)")
+                    .unwrap()
+                    .progress_chars("=>-");
+                pb.set_style(style);
+                pb.set_prefix(format!("{}", m.name));
+                Some(pb)
+            } else { None };
+            handles.push(std::thread::spawn(move || {
+                if let Err(e) = download_one_model(&client2, &models_dir2, &m, pb_opt) {
+                    crate::elog!("Error: {:#}", e);
+                }
+            }));
         }
+        for h in handles { let _ = h.join(); }
+        i = end;
     }
+
+    // Drop MultiProgress after threads are joined; bars finish naturally.
+    drop(mp_opt);
+
     Ok(())
 }
 
 /// Download a single model entry into the given models directory, verifying SHA-256 when available.
-fn download_one_model(client: &Client, models_dir: &Path, entry: &ModelEntry) -> Result<()> {
+fn download_one_model(client: &Client, models_dir: &Path, entry: &ModelEntry, pb: Option<indicatif::ProgressBar>) -> Result<()> {
     let final_path = models_dir.join(format!("ggml-{}.bin", entry.name));
 
     // If the model already exists, verify against online metadata
@@ -637,9 +676,11 @@ fn download_one_model(client: &Client, models_dir: &Path, entry: &ModelEntry) ->
             .with_context(|| format!("Failed to create {}", tmp_path.display()))?,
     );
 
-    // Set up progress bar if interactive and we know size
+    // Set up progress bar: use provided one if present; otherwise create if interactive and we know size
     let show_progress = !crate::is_quiet() && !crate::is_no_progress() && atty::is(Stream::Stderr) && entry.size > 0;
-    let pb_opt = if show_progress {
+    let pb_opt = if let Some(p) = pb {
+        Some(p)
+    } else if show_progress {
         let pb = ProgressBar::new(entry.size);
         let style = ProgressStyle::with_template("Downloading {prefix} ({total_bytes}) [{bar:40.cyan/blue}] {bytes}/{total_bytes} ({percent}%)")
             .unwrap()
@@ -804,18 +845,18 @@ pub fn update_local_models() -> Result<()> {
                         wlog!("Failed hashing {}: {}. Re-downloading.", fname, e);
                     }
                 }
-                download_one_model(&client, models_dir, remote)?;
+                download_one_model(&client, models_dir, remote, None)?;
             } else if remote.size > 0 {
                 match std::fs::metadata(&path) {
                     Ok(md) => {
                         if qlog_size_comparison(&fname, md.len(), remote.size) {
                             continue;
                         }
-                        download_one_model(&client, models_dir, remote)?;
+                        download_one_model(&client, models_dir, remote, None)?;
                     }
                     Err(e) => {
                         wlog!("Stat failed for {}: {}. Updating...", fname, e);
-                        download_one_model(&client, models_dir, remote)?;
+                        download_one_model(&client, models_dir, remote, None)?;
                     }
                 }
             } else {
@@ -883,7 +924,7 @@ pub fn ensure_model_available_noninteractive(model_name: &str) -> Result<std::pa
     // Prefer fetching metadata to construct a proper ModelEntry
     let models = fetch_all_models(&client)?;
     if let Some(entry) = models.into_iter().find(|m| m.name == model_name) {
-        download_one_model(&client, models_dir, &entry)?;
+        download_one_model(&client, models_dir, &entry, None)?;
         return Ok(models_dir.join(format!("ggml-{}.bin", entry.name)));
     }
     Err(anyhow!(