diff --git a/base-convert/crates/base-convert/src/hub.rs b/base-convert/crates/base-convert/src/hub.rs
index 69b5420..a25d558 100644
--- a/base-convert/crates/base-convert/src/hub.rs
+++ b/base-convert/crates/base-convert/src/hub.rs
@@ -313,6 +313,7 @@ fn auto_pull_and_resolve(
 
     cmd_pull(PullArgs {
         id: pull_id.clone(),
+        variant: None,
         profile: None,
         target,
         revision: "main".to_string(),
@@ -568,7 +569,39 @@ const RUNTIME_COMMANDS: [&str; 6] = [
     "profile",
 ];
 
-pub fn cmd_pull(args: PullArgs) -> Result<()> {
+/// Fold a `:variant` suffix on the id (or `--variant`) into the pull: the id
+/// loses the suffix and the variant's quant picks `--target`, the same mapping
+/// the run commands use when they auto-pull. Without this the suffix went to
+/// the Hub as part of the repo name (`basecompute/Qwen3-8B:default-q8@main` →
+/// 404). The inline form wins over the flag, as it does for `serve`/`chat`.
+fn apply_pull_variant(args: &mut PullArgs) -> Result<()> {
+    let (id, inline) = match args.id.split_once(':') {
+        Some((id, v)) => (id.to_string(), Some(v.to_string())),
+        None => (args.id.clone(), None),
+    };
+    if inline.as_deref() == Some("") {
+        bail!(
+            "empty variant after ':' in `{}` — name a variant (e.g. `{id}:default-q8`) \
+             or drop the colon",
+            args.id
+        );
+    }
+    if let Some(v) = inline.or_else(|| args.variant.clone()) {
+        if v.is_empty() {
+            bail!("`--variant` value is empty");
+        }
+        let quant = base_hub::registry::quant_bits(&v)
+            .map(str::to_string)
+            .unwrap_or_else(|| quant_tag(&v));
+        args.target = target_from_quant(&quant);
+        args.variant = Some(v);
+    }
+    args.id = id;
+    Ok(())
+}
+
+pub fn cmd_pull(mut args: PullArgs) -> Result<()> {
+    apply_pull_variant(&mut args)?;
     let reg = MergedRegistry::load()?;
     let root = reg.root.clone();
     let want = quant_token(&args);
@@ -1444,6 +1477,7 @@ mod tests {
         let root = tmp.path().join("cache");
         let args = PullArgs {
             id: "basecompute/m".into(),
+            variant: None,
             profile: None,
             target: TargetScheme::BaseQ4,
             revision: "main".into(),
@@ -1488,6 +1522,7 @@ mod tests {
         let root = tmp.path().join("cache");
         let args = PullArgs {
             id: "basecompute/m".into(),
+            variant: None,
             profile: None,
             target: TargetScheme::BaseQ8,
             revision: "main".into(),
@@ -1576,6 +1611,7 @@ mod tests {
 
         let args = PullArgs {
             id: "basecompute/m".into(),
+            variant: None,
             profile: None,
             target: TargetScheme::BaseQ4,
             revision: "main".into(),
@@ -1645,6 +1681,56 @@ mod tests {
         );
     }
 
+    #[test]
+    fn apply_pull_variant_splits_inline_and_flag() {
+        let mk = |id: &str, variant: Option<&str>| PullArgs {
+            id: id.into(),
+            variant: variant.map(Into::into),
+            profile: None,
+            target: TargetScheme::BaseQ4,
+            revision: "main".into(),
+            force: false,
+            dry_run: false,
+        };
+
+        // Inline `:variant` comes off the id and picks the quant (#70).
+        let mut a = mk("basecompute/Qwen3-8B:default-q8", None);
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.id, "basecompute/Qwen3-8B");
+        assert_eq!(a.target, TargetScheme::BaseQ8);
+        assert_eq!(quant_token(&a), "q8");
+
+        // `--variant` does the same.
+        let mut a = mk("basecompute/Qwen3-8B", Some("default-q8"));
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.id, "basecompute/Qwen3-8B");
+        assert_eq!(a.target, TargetScheme::BaseQ8);
+
+        // Inline wins over the flag.
+        let mut a = mk("basecompute/Qwen3-8B:default-q8", Some("default-q4"));
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.target, TargetScheme::BaseQ8);
+
+        // Backend-native names map by their bit width.
+        let mut a = mk("basecompute/Qwen3-8B:cuda-q4mix", None);
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.target, TargetScheme::BaseQ4);
+        let mut a = mk("basecompute/gpt-oss-20b:default-mxfp4", None);
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.target, TargetScheme::Mxfp4);
+
+        // No variant: id and `--target` untouched.
+        let mut a = mk("basecompute/Qwen3-8B", None);
+        a.target = TargetScheme::BaseQ6;
+        apply_pull_variant(&mut a).unwrap();
+        assert_eq!(a.id, "basecompute/Qwen3-8B");
+        assert_eq!(a.target, TargetScheme::BaseQ6);
+
+        // A trailing colon or an empty flag is an error, not a silent q4.
+        assert!(apply_pull_variant(&mut mk("basecompute/Qwen3-8B:", None)).is_err());
+        assert!(apply_pull_variant(&mut mk("basecompute/Qwen3-8B", Some(""))).is_err());
+    }
+
     #[test]
     fn target_from_quant_maps_known_tags() {
         assert!(matches!(target_from_quant("q8"), TargetScheme::BaseQ8));
@@ -1942,6 +2028,7 @@ mod tests {
     fn choose_profile_uses_bundled_for_q4_q6_and_q8() {
         let mk = |t| PullArgs {
             id: "x/y".into(),
+            variant: None,
             profile: None,
             target: t,
             revision: "main".into(),
diff --git a/base-convert/crates/base-convert/src/main.rs b/base-convert/crates/base-convert/src/main.rs
index 91fdcd6..ff2dcb0 100644
--- a/base-convert/crates/base-convert/src/main.rs
+++ b/base-convert/crates/base-convert/src/main.rs
@@ -365,8 +365,13 @@ enum AwqMode {
 #[derive(Parser, Debug)]
 struct PullArgs {
     /// Model id: `basecompute/<name>` (pre-converted, from the catalog) or
-    /// `org/model` (a raw HF repo, downloaded + converted locally).
+    /// `org/model` (a raw HF repo, downloaded + converted locally). Append
+    /// `:<variant>` (e.g. `:default-q8`) to pick a quant build.
     id: String,
+    /// Quant build to pull, e.g. `default-q8`; same as `<id>:default-q8`
+    /// (an inline `:variant` wins). Overrides `--target`.
+    #[arg(long)]
+    variant: Option<String>,
     /// Override the auto-selected quant profile (path to a profile JSON).
     #[arg(long)]
     profile: Option<PathBuf>,