diff --git a/Sources/Tokenizers/Decoder.swift b/Sources/Tokenizers/Decoder.swift
index 49021645..80d4e302 100644
--- a/Sources/Tokenizers/Decoder.swift
+++ b/Sources/Tokenizers/Decoder.swift
@@ -199,6 +199,10 @@ class ByteFallbackDecoder: Decoder {
                 newTokens.append(token)
             }
         }
+        if !byteTokens.isEmpty {
+            let codeUnits = byteTokens.map { UTF8.CodeUnit($0) }
+            newTokens.append(String(decoding: codeUnits, as: UTF8.self))
+        }
         return newTokens
     }
 }
diff --git a/Tests/TokenizersTests/ByteFallbackEndOfInputTests.swift b/Tests/TokenizersTests/ByteFallbackEndOfInputTests.swift
new file mode 100644
index 00000000..93a207f3
--- /dev/null
+++ b/Tests/TokenizersTests/ByteFallbackEndOfInputTests.swift
@@ -0,0 +1,76 @@
+import Foundation
+import Hub
+import Testing
+
+@testable import Tokenizers
+
+@Suite("Byte fallback end-of-input")
+struct ByteFallbackEndOfInputTests {
+    @Test("Terminal UTF-8 bytes decode without a following ordinary token")
+    func terminalEmoji() {
+        let decoder = ByteFallbackDecoder(config: Config(["type": "ByteFallback"] as [NSString: Any]))
+        #expect(decoder.decode(tokens: ["<0xF0>", "<0x9F>", "<0x98>", "<0x80>"]) == ["😀"])
+    }
+
+    @Test("Interior and terminal byte runs retain their order")
+    func mixedRuns() {
+        let decoder = ByteFallbackDecoder(config: Config(["type": "ByteFallback"] as [NSString: Any]))
+        #expect(decoder.decode(tokens: ["a", "<0x62>", "c", "<0x20>"]) == ["a", "b", "c", " "])
+        #expect(decoder.decode(tokens: []) == [])
+        #expect(decoder.decode(tokens: ["ordinary"]) == ["ordinary"])
+    }
+
+    private func pipeline() -> Config {
+        Config(
+            [
+                "type": "Sequence",
+                "decoders": [
+                    ["type": "Replace", "pattern": ["String": "▁"], "content": " "],
+                    ["type": "ByteFallback"],
+                    ["type": "Fuse"],
+                    ["type": "Strip", "content": " ", "start": 0, "stop": 1],
+                ],
+            ] as [NSString: Any])
+    }
+
+    @Test("A terminal byte-spelled space still passes through downstream Fuse and Strip")
+    func downstreamCleanup() throws {
+        let decoder = try #require(try DecoderFactory.fromConfig(config: pipeline()))
+        #expect(decoder.decode(tokens: ["<0x20>"]) == [""])
+        #expect(decoder.decode(tokens: ["a", "<0x20>"]) == ["a"])
+        #expect(decoder.decode(tokens: ["▁"]) == [""])
+    }
+
+    @Test("PreTrainedTokenizer preserves terminal bytes, special-token filtering and configured cleanup")
+    func configuredTokenizer() throws {
+        let config = Config(
+            [
+                "tokenizer_class": "PreTrainedTokenizerFast",
+                "unk_token": "<unk>",
+                "eos_token": "</s>",
+                "clean_up_tokenization_spaces": false,
+            ] as [NSString: Any])
+        let data = Config(
+            [
+                "added_tokens": [
+                    ["id": 0, "content": "<unk>", "special": true],
+                    ["id": 1, "content": "</s>", "special": true],
+                ],
+                "decoder": pipeline(),
+                "model": [
+                    "type": "BPE", "unk_token": "<unk>", "byte_fallback": true,
+                    "vocab": [
+                        "<unk>": 0, "</s>": 1, "a": 2, "▁": 3,
+                        "<0x20>": 4, "<0xF0>": 5, "<0x9F>": 6, "<0x98>": 7, "<0x80>": 8,
+                    ],
+                    "merges": [String](),
+                ],
+            ] as [NSString: Any])
+        let tokenizer = try PreTrainedTokenizer(tokenizerConfig: config, tokenizerData: data)
+        #expect(tokenizer.decode(tokens: [5, 6, 7, 8]) == "😀")
+        #expect(tokenizer.decode(tokens: [4]) == tokenizer.decode(tokens: [3]))
+        #expect(tokenizer.decode(tokens: [2, 4]) == "a")
+        #expect(tokenizer.decode(tokens: [1, 5, 6, 7, 8], skipSpecialTokens: true) == "😀")
+        #expect(tokenizer.decode(tokens: [1, 5, 6, 7, 8], skipSpecialTokens: false) == "</s>😀")
+    }
+}