diff --git a/open-sse/services/model.ts b/open-sse/services/model.ts
index d3b1210dcf..c0fd9582d6 100644
--- a/open-sse/services/model.ts
+++ b/open-sse/services/model.ts
@@ -496,6 +496,23 @@ export async function getModelInfoCore(
   // Get aliases (from object or function)
   const aliases = typeof aliasesOrGetter === "function" ? await aliasesOrGetter() : aliasesOrGetter;
 
+  // Local alias map (user-provided 2nd arg) wins over all cross-proxy /
+  // provider inference paths. When the alias target is a slashful string like
+  // "openai/gpt-4o", parse it directly as <provider>/<model> and return
+  // immediately — before shouldTreatAsExactModelId() or cross-proxy inference
+  // can misclassify the target (e.g. because bazaarlink catalogs it verbatim).
+  if (aliases && parsed.model && typeof aliases[parsed.model] === "string") {
+    const directTarget = aliases[parsed.model];
+    const slashIdx = directTarget.indexOf("/");
+    if (slashIdx !== -1) {
+      const providerPart = directTarget.slice(0, slashIdx);
+      const modelPart = directTarget.slice(slashIdx + 1);
+      const provider = resolveProviderAlias(providerPart);
+      const canonicalModel = resolveProviderModelAlias(provider, modelPart);
+      return { provider, model: canonicalModel, extendedContext };
+    }
+  }
+
   // Resolve exact alias
   const resolved = resolveModelAliasTarget(parsed.model, aliases);
   if (resolved?.provider) {
diff --git a/tests/unit/local-aliases-precedence.test.ts b/tests/unit/local-aliases-precedence.test.ts
new file mode 100644
index 0000000000..54df2e1b8f
--- /dev/null
+++ b/tests/unit/local-aliases-precedence.test.ts
@@ -0,0 +1,74 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { getModelInfoCore } from "../../open-sse/services/model.ts";
+
+// These tests cover the early-return path added in this PR (open-sse/services/model.ts):
+// when a local alias map provides a slashful string target (e.g. "openai/gpt-4o"),
+// the alias target is parsed directly as <provider>/<model> and returned
+// immediately, before shouldTreatAsExactModelId() / cross-proxy inference can
+// misclassify the target.
+
+test("local alias with slashful string target returns provider/model directly", async () => {
+  const result = await getModelInfoCore("my-alias", {
+    "my-alias": "openai/gpt-4o",
+  });
+  assert.deepEqual(result, {
+    provider: "openai",
+    model: "gpt-4o",
+    extendedContext: false,
+  });
+});
+
+test("local alias early-return preserves extendedContext from [1m] suffix on input", async () => {
+  const result = await getModelInfoCore("my-alias[1m]", {
+    "my-alias": "openai/gpt-4o",
+  });
+  assert.deepEqual(result, {
+    provider: "openai",
+    model: "gpt-4o",
+    extendedContext: true,
+  });
+});
+
+test("local alias early-return preserves the slashful target's model part verbatim (no cross-proxy normalization)", async () => {
+  // The early-return is intentionally narrower than the object-target path:
+  // it only applies provider-scoped alias resolution, not the global
+  // cross-proxy normalization (CROSS_PROXY_MODEL_ALIASES). This mirrors the
+  // PR's intent — "user-provided 2nd arg wins, parsed directly" — and avoids
+  // double-resolving aliases the user already pinned to a canonical pair.
+  const result = await getModelInfoCore("sf-qwen", {
+    "sf-qwen": "siliconflow/qwen3-coder:480b",
+  });
+  assert.deepEqual(result, {
+    provider: "siliconflow",
+    model: "qwen3-coder:480b",
+    extendedContext: false,
+  });
+});
+
+test("local alias with non-slashful string target falls through to standard alias resolution", async () => {
+  // No "/" in the target → early-return guard doesn't fire; existing
+  // resolveModelAliasTarget path handles it. We assert that the legacy path
+  // still works and isn't shadowed by the new shortcut.
+  const result = await getModelInfoCore("legacy-alias", {
+    "legacy-alias": "gpt-oss:120b",
+  });
+  // The legacy path resolves through cross-proxy / provider inference;
+  // we only assert that the early-return did NOT trip (model is normalized,
+  // not returned verbatim with provider=null,model="gpt-oss:120b").
+  assert.notEqual(result.model, "gpt-oss:120b");
+});
+
+test("local alias with object target (not string) falls through to standard resolution", async () => {
+  // The early-return guard checks `typeof aliases[parsed.model] === "string"`.
+  // Object targets must keep going through resolveModelAliasTarget.
+  const result = await getModelInfoCore("sf-qwen-obj", {
+    "sf-qwen-obj": { provider: "siliconflow", model: "qwen3-coder:480b" },
+  });
+  assert.deepEqual(result, {
+    provider: "siliconflow",
+    model: "Qwen/Qwen3-Coder-480B-A35B-Instruct",
+    extendedContext: false,
+  });
+});