Add Apertus tool parser

mlx_lm/tokenizer_utils.py

@@ -253,6 +253,9 @@ def make_byte_decoder(cls):
 def _infer_thinking(tokenizer):
     vocab = tokenizer.get_vocab()
     THINK_TOKENS = [
+        # Apertus has <think>/</think> in its vocabulary but never emits them,
+        # so its own markers must be checked first.
+        ("<|inner_prefix|>", "<|inner_suffix|>"),
         ("<think>", "</think>"),
         ("<longcat_think>", "</longcat_think>"),
     ]
@@ -566,6 +569,8 @@ def _infer_tool_parser(chat_template):
         return "qwen3_coder"
     elif "<|tool_calls_section_begin|>" in chat_template:
         return "kimi_k2"
+    elif "<|tools_prefix|>" in chat_template:
+        return "apertus"
     elif "[TOOL_CALLS]" in chat_template:
         return "mistral"
     elif "<tool_call>" in chat_template and "tool_call.name" in chat_template:

mlx_lm/tool_parsers/apertus.py

@@ -0,0 +1,35 @@
+# Copyright © 2026 Apple Inc.
+"""
+Modified from:
+https://github.com/vllm-project/vllm/blob/main/vllm/tool_parsers/apertus_tool_parser.py
+"""
+
+import json
+from typing import Any
+
+tool_call_start = "<|tools_prefix|>"
+tool_call_end = "<|tools_suffix|>"
+
+
+def parse_tool_call(text: str, tools: Any | None = None):
+    # Apertus emits an array of single key objects mapping the function name to
+    # its arguments, e.g.
+    #   [{"get_weather": {"location": "London"}}, {"get_time": {}}]
+    calls = json.loads(text)
+    if not isinstance(calls, list):
+        calls = [calls]
+
+    tool_calls = []
+    for call in calls:
+        if not isinstance(call, dict) or not call:
+            continue
+        name, arguments = next(iter(call.items()))
+        # A call with no arguments can come back as null, but clients expect an
+        # object.
+        if arguments is None:
+            arguments = {}
+        tool_calls.append(dict(name=name, arguments=arguments))
+
+    if not tool_calls:
+        raise ValueError(f"Could not parse tool call from: {text}")
+    return tool_calls

tests/test_tokenizers.py

@@ -10,6 +10,7 @@
     NaiveStreamingDetokenizer,
     SPMStreamingDetokenizer,
     TokenizerWrapper,
+    _infer_thinking,
 )
 from mlx_lm.utils import load_tokenizer
 
@@ -110,6 +111,35 @@ def test_thinking(self):
         self.assertIsNone(tokenizer.think_start_id)
         self.assertIsNone(tokenizer.think_end_id)
 
+    def test_thinking_marker_precedence(self):
+        # Apertus carries unused <think> tokens alongside the markers it
+        # actually emits, so its own markers have to win.
+        class Tokenizer:
+            def __init__(self, vocab):
+                self._vocab = vocab
+
+            def get_vocab(self):
+                return self._vocab
+
+        apertus_vocab = {
+            "<|inner_prefix|>": 32,
+            "<|inner_suffix|>": 33,
+            "<think>": 69,
+            "</think>": 70,
+        }
+        think_start, think_end, start_ids, end_ids = _infer_thinking(
+            Tokenizer(apertus_vocab)
+        )
+        self.assertEqual(think_start, "<|inner_prefix|>")
+        self.assertEqual(think_end, "<|inner_suffix|>")
+        self.assertEqual(start_ids, (32,))
+        self.assertEqual(end_ids, (33,))
+
+        think_start, _, _, _ = _infer_thinking(Tokenizer({"<think>": 1, "</think>": 2}))
+        self.assertEqual(think_start, "<think>")
+
+        self.assertEqual(_infer_thinking(Tokenizer({})), (None, None, None, None))
+
     def test_find_token(self):
         # Check that _find returns a valid index when
         # searching for a think token in short system prompts

tests/test_tool_parsing.py

@@ -2,6 +2,7 @@
 from pathlib import Path
 
 from mlx_lm.tool_parsers import (
+    apertus,
     function_gemma,
     gemma4,
     glm47,
@@ -313,6 +314,61 @@ def test_kimi_k2(self):
         ]
         self.assertEqual(tool_calls, expected)
 
+    def test_apertus(self):
+        # Single tool call
+        test_case = '[{"multiply": {"a": 12234585, "b": 48838483920}}]'
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        expected = [
+            {"name": "multiply", "arguments": {"a": 12234585, "b": 48838483920}}
+        ]
+        self.assertEqual(tool_calls, expected)
+
+        # Multiple tool calls
+        test_case = (
+            '[{"get_weather": {"location": "London"}}, '
+            '{"get_time": {"location": "London"}}]'
+        )
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        expected = [
+            {"name": "get_weather", "arguments": {"location": "London"}},
+            {"name": "get_time", "arguments": {"location": "London"}},
+        ]
+        self.assertEqual(tool_calls, expected)
+
+        # Nested arguments
+        test_case = (
+            '[{"complex_function": {"nested": {"inner": "value"}, "list": ["a", "b"]}}]'
+        )
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        expected = [
+            {
+                "name": "complex_function",
+                "arguments": {"nested": {"inner": "value"}, "list": ["a", "b"]},
+            }
+        ]
+        self.assertEqual(tool_calls, expected)
+
+        # Null arguments are normalized so clients receive an empty object
+        test_case = '[{"get_time": null}]'
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        self.assertEqual(tool_calls, [{"name": "get_time", "arguments": {}}])
+
+        # A bare object rather than an array
+        test_case = '{"get_time": {}}'
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        self.assertEqual(tool_calls, [{"name": "get_time", "arguments": {}}])
+
+        # Entries which cannot be a tool call are skipped
+        test_case = '[{"get_time": {}}, {}, "junk"]'
+        tool_calls = apertus.parse_tool_call(test_case, None)
+        self.assertEqual(tool_calls, [{"name": "get_time", "arguments": {}}])
+
+        # Nothing parseable raises, as does a call truncated mid generation
+        with self.assertRaises(ValueError):
+            apertus.parse_tool_call("[]", None)
+        with self.assertRaises(ValueError):
+            apertus.parse_tool_call('[{"get_weather": {"location": "London"}', None)
+
     def test_minimax_m2(self):
         test_case = (
             '<invoke name="search">\n'