nanbeige-mlx 0.2.0__tar.gz

nanbeige_mlx-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Jishnu Venugopal
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

nanbeige_mlx-0.2.0/LICENSE.apache

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

nanbeige_mlx-0.2.0/PKG-INFO

@@ -0,0 +1,104 @@
+Metadata-Version: 2.4
+Name: nanbeige-mlx
+Version: 0.2.0
+Summary: MLX port of the Nanbeige4.2-3B Looped Transformer for Apple Silicon — model definition, HF->MLX conversion, and publish helpers.
+Author: Jishnu Venugopal
+License: MIT
+Project-URL: Homepage, https://github.com/jishnuvenugopal/nanbeige-mlx
+Project-URL: Repository, https://github.com/jishnuvenugopal/nanbeige-mlx
+Project-URL: Issues, https://github.com/jishnuvenugopal/nanbeige-mlx/issues
+Project-URL: Changelog, https://github.com/jishnuvenugopal/nanbeige-mlx/blob/main/CHANGELOG.md
+Project-URL: Evaluation harness, https://github.com/jishnuvenugopal/nanbeige-mlx-eval
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: MacOS
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: LICENSE.apache
+Requires-Dist: mlx<0.34,>=0.32
+Requires-Dist: mlx-lm<0.33,>=0.31
+Requires-Dist: huggingface_hub>=1.20
+Requires-Dist: transformers>=4.40
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Dynamic: license-file
+
+# nanbeige-mlx
+
+An MLX port of the **Nanbeige4.2-3B *Looped Transformer*** for Apple Silicon,
+plus the HF→MLX conversion and publishing helpers. The model definition in
+[`nanbeige_mlx/model.py`](nanbeige_mlx/model.py) is the **single source of truth**
+for the port — it is also copied verbatim into each converted weight directory
+as the `model_file` mlx-lm loads.
+
+This is an independent project; it is not affiliated with or endorsed by the
+Nanbeige team.
+
+## What's here
+
+| module | purpose |
+|---|---|
+| `nanbeige_mlx/model.py` | the port — also the shipped `model_file` |
+| `nanbeige_mlx/convert.py` | HF → MLX quant (non-mutating staging, tokenizer verify) |
+| `nanbeige_mlx/upload.py` | model-card + LICENSE + NOTICE + (opt-in) HF upload |
+| `nanbeige_mlx/pull.py` | `pull("4bit")` → local path for `mlx_lm.load` |
+
+## Load a published quant (one line)
+
+```python
+from nanbeige_mlx import pull
+import mlx_lm
+model, tok = mlx_lm.load(pull("4bit"))
+```
+
+## Convert from the BF16 checkpoint yourself
+
+```bash
+nanbeige-mlx-convert --src /path/to/Nanbeige4.2-3B --out ./nanbeige-mlx-4bit --bits 4
+```
+
+The source directory is never mutated; the tokenizer round-trip is asserted.
+
+## The 44-slot KV cache (a real cost of the looped design)
+
+The looped architecture needs `num_loops * num_hidden_layers = 44` KV slots.
+At full context that is 44 × 8 KV-heads × 128 dim × 2 (K+V) × 262 144 positions
+× 2 bytes ≈ **47 GB** — unreachable on a 16 GB machine. Because this model
+supplies `make_cache`, mlx-lm's `--max-kv-size` knob is inert; use `--kv-bits`
+to reduce KV precision instead. This is a quantifiable cost of the looped
+design, not a bug — see `make_cache()`'s docstring.
+
+## Dependencies
+
+Upper-bounded to `mlx>=0.32,<0.34` and `mlx-lm>=0.31,<0.33`. The bounds are
+deliberate: `model.py` uses `mlx_lm.models.base` / `.cache` internals and is
+copied verbatim into every published weight repo as the `model_file`, where it
+is frozen — a user who already downloaded a quant cannot receive a patch.
+Relaxing the bound is a decision to make after testing against a new mlx-lm.
+
+## Fidelity status
+
+This port is validated by behaviour, not bit-parity: per-layer arithmetic
+agrees with the reference's own `NanbeigeDecoderLayer` to fp32 precision, the
+44-slot loop-aware KV cache passes a prefill-vs-incremental-decode equality
+test, and the port's two code paths are bit-identical. End-to-end next-token
+logit cosine against the HF reference is 0.847 (top-1 agreement 83%), lower
+than a faithful port should give; six candidate causes have been eliminated by
+measurement. The full record — including every falsified hypothesis — is in
+the evaluation harness repo's
+[`docs/investigation-log.md`](https://github.com/jishnuvenugopal/nanbeige-mlx-eval/blob/main/docs/investigation-log.md).
+Behaviour on the bilingual agentic suite is unaffected (26–28/30 across
+4/6/8-bit).
+
+## License
+
+MIT for the code in this package. The Nanbeige model weights are governed by
+the upstream Apache-2.0 license; convert and redistribute them per that license.

nanbeige_mlx-0.2.0/README.md

@@ -0,0 +1,71 @@
+# nanbeige-mlx
+
+An MLX port of the **Nanbeige4.2-3B *Looped Transformer*** for Apple Silicon,
+plus the HF→MLX conversion and publishing helpers. The model definition in
+[`nanbeige_mlx/model.py`](nanbeige_mlx/model.py) is the **single source of truth**
+for the port — it is also copied verbatim into each converted weight directory
+as the `model_file` mlx-lm loads.
+
+This is an independent project; it is not affiliated with or endorsed by the
+Nanbeige team.
+
+## What's here
+
+| module | purpose |
+|---|---|
+| `nanbeige_mlx/model.py` | the port — also the shipped `model_file` |
+| `nanbeige_mlx/convert.py` | HF → MLX quant (non-mutating staging, tokenizer verify) |
+| `nanbeige_mlx/upload.py` | model-card + LICENSE + NOTICE + (opt-in) HF upload |
+| `nanbeige_mlx/pull.py` | `pull("4bit")` → local path for `mlx_lm.load` |
+
+## Load a published quant (one line)
+
+```python
+from nanbeige_mlx import pull
+import mlx_lm
+model, tok = mlx_lm.load(pull("4bit"))
+```
+
+## Convert from the BF16 checkpoint yourself
+
+```bash
+nanbeige-mlx-convert --src /path/to/Nanbeige4.2-3B --out ./nanbeige-mlx-4bit --bits 4
+```
+
+The source directory is never mutated; the tokenizer round-trip is asserted.
+
+## The 44-slot KV cache (a real cost of the looped design)
+
+The looped architecture needs `num_loops * num_hidden_layers = 44` KV slots.
+At full context that is 44 × 8 KV-heads × 128 dim × 2 (K+V) × 262 144 positions
+× 2 bytes ≈ **47 GB** — unreachable on a 16 GB machine. Because this model
+supplies `make_cache`, mlx-lm's `--max-kv-size` knob is inert; use `--kv-bits`
+to reduce KV precision instead. This is a quantifiable cost of the looped
+design, not a bug — see `make_cache()`'s docstring.
+
+## Dependencies
+
+Upper-bounded to `mlx>=0.32,<0.34` and `mlx-lm>=0.31,<0.33`. The bounds are
+deliberate: `model.py` uses `mlx_lm.models.base` / `.cache` internals and is
+copied verbatim into every published weight repo as the `model_file`, where it
+is frozen — a user who already downloaded a quant cannot receive a patch.
+Relaxing the bound is a decision to make after testing against a new mlx-lm.
+
+## Fidelity status
+
+This port is validated by behaviour, not bit-parity: per-layer arithmetic
+agrees with the reference's own `NanbeigeDecoderLayer` to fp32 precision, the
+44-slot loop-aware KV cache passes a prefill-vs-incremental-decode equality
+test, and the port's two code paths are bit-identical. End-to-end next-token
+logit cosine against the HF reference is 0.847 (top-1 agreement 83%), lower
+than a faithful port should give; six candidate causes have been eliminated by
+measurement. The full record — including every falsified hypothesis — is in
+the evaluation harness repo's
+[`docs/investigation-log.md`](https://github.com/jishnuvenugopal/nanbeige-mlx-eval/blob/main/docs/investigation-log.md).
+Behaviour on the bilingual agentic suite is unaffected (26–28/30 across
+4/6/8-bit).
+
+## License
+
+MIT for the code in this package. The Nanbeige model weights are governed by
+the upstream Apache-2.0 license; convert and redistribute them per that license.

nanbeige_mlx-0.2.0/nanbeige_mlx/__init__.py

@@ -0,0 +1,22 @@
+"""nanbeige-mlx: an MLX port of the Nanbeige4.2-3B Looped Transformer.
+
+The model definition lives in :mod:`nanbeige_mlx.model` and is the single
+source of truth — it is also copied verbatim into each converted weight
+directory as the ``model_file`` mlx-lm loads. Conversion (HF -> MLX quant) is
+in :mod:`nanbeige_mlx.convert`; publishing helpers in
+:mod:`nanbeige_mlx.upload` and :mod:`nanbeige_mlx.pull`.
+"""
+
+from __future__ import annotations
+
+__version__ = "0.2.0"
+
+
+def __getattr__(name: str):  # pragma: no cover - thin convenience export
+    # Lazy re-export so ``from nanbeige_mlx import pull`` works without importing
+    # huggingface_hub at package-import time.
+    if name == "pull":
+        from .pull import pull
+
+        return pull
+    raise AttributeError(f"module 'nanbeige_mlx' has no attribute {name!r}")

nanbeige_mlx-0.2.0/nanbeige_mlx/convert.py

@@ -0,0 +1,235 @@
+"""Convert Nanbeige4.2-3B HF weights into MLX quants.
+
+Uses mlx-lm's ``model_file`` hook: the MLX port (``model.py``) is dropped into a
+*staged copy* of the HF repo and referenced from that copy's ``config.json``, so
+the standard ``mlx_lm.convert`` pipeline (load -> quantize -> save) works without
+a registry entry. ``auto_map`` is stripped from the *output* config because the
+MLX side loads via ``model_file``; it is left intact in the source so the HF
+parity reference can still load the remote modeling code.
+
+Two correctness properties (P3.3, P3.4 in docs/investigation-log.md):
+
+* **Never mutate the source.** ``prepare_source`` stages a conversion-ready copy
+  in a temp dir, symlinks the (large) safetensors shards instead of duplicating
+  them, and writes the ``model_file`` config into the *copy*. Pointing ``--src``
+  at an HF cache snapshot is therefore safe — the snapshot's hashes are
+  untouched. ``to_mlx`` calls ``prepare_source`` itself and cleans up the stage,
+  so ``convert`` is one step and idempotent from a pristine HF download.
+* **Preserve tokenizer files and verify them.** mlx-lm's convert writes a
+  trimmed ``tokenizer_config.json`` and drops ``added_tokens.json`` /
+  ``special_tokens_map.json`` / ``tokenizer.model``. We copy those through,
+  drop transient init keys, restore ``model_max_length``, and assert the
+  tokenizer round-trips the EOS token before declaring success.
+"""
+
+from __future__ import annotations
+
+import json
+import shutil
+import tempfile
+from pathlib import Path
+
+from mlx_lm.convert import convert
+
+# The single source of truth for the port. Copied verbatim into each staged
+# (and therefore each converted) weight repo as the shipped ``model_file``.
+PORT_FILE = Path(__file__).resolve().parent / "model.py"
+
+# Tokenizer / template files mlx-lm's convert trims that we must carry through.
+CARRY = [
+    "added_tokens.json",
+    "special_tokens_map.json",
+    "tokenizer.model",
+    "chat_template.jinja",
+]
+
+# Transient keys mlx-lm leaks into tokenizer_config.json from
+# tokenizer.init_kwargs; they don't belong in a published config.
+_TRANSIENT_TOKENIZER_KEYS = ("is_local", "local_files_only", "backend")
+
+# HF ``trust_remote_code`` files that the source repo carries for the PyTorch
+# reference path. They must NOT ship in an MLX repo (which loads via
+# ``model_file``), so remove them from the output after convert.
+_STRIP_FROM_OUTPUT = {
+    "configuration_nanbeige.py",
+    "modeling_nanbeige.py",
+}
+
+# Nanbeige4.2-3B supports 256K context; mlx-lm's convert writes the int64
+# sentinel instead of this concrete value. Restore it.
+MODEL_MAX_LENGTH = 262144
+
+# EOS token the model must be able to emit to ever stop generating.
+EOS_TOKEN = "<|im_end|>"
+EOS_TOKEN_ID = 166101
+
+
+def prepare_source(
+    src_dir: str | Path,
+    model_file: str | Path = PORT_FILE,
+    workdir: str | Path | None = None,
+) -> Path:
+    """Stage a conversion-ready copy of ``src_dir``. **Never writes into ``src_dir``**.
+
+    Non-safetensor files are copied; ``*.safetensors`` shards are *symlinked*
+    (not duplicated — the BF16 checkpoint is ~8 GB). The port is copied in as
+    ``nanbeige.py`` and ``config.json["model_file"]`` is set on the *copy*.
+
+    Returns the stage directory. The caller is responsible for cleanup if they
+    don't go through :func:`to_mlx` (which handles it).
+    """
+    src = Path(src_dir)
+    if not src.is_dir():
+        raise FileNotFoundError(f"source repo not found: {src}")
+
+    stage = Path(workdir) if workdir else Path(
+        tempfile.mkdtemp(prefix="nanbeige-stage-")
+    )
+    stage.mkdir(parents=True, exist_ok=True)
+
+    # Copy small files verbatim.
+    for f in src.iterdir():
+        if f.is_file() and f.suffix != ".safetensors":
+            shutil.copy2(f, stage / f.name)
+
+    # Symlink the weight shards — don't duplicate ~8 GB.
+    for f in src.glob("*.safetensors"):
+        (stage / f.name).symlink_to(f.resolve())
+
+    # Drop in the port and wire up model_file on the staged config copy.
+    shutil.copy2(model_file, stage / "nanbeige.py")
+    cfg_path = stage / "config.json"
+    cfg = json.loads(cfg_path.read_text(encoding="utf-8"))
+    cfg["model_file"] = "nanbeige.py"
+    cfg_path.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    return stage
+
+
+def _finalize(src: Path, out: Path) -> None:
+    """Carry tokenizer files through, scrub transient keys, restore max length.
+
+    Idempotent: only copies a CARRY file when the output is missing it.
+    """
+    for name in CARRY:
+        if (src / name).exists() and not (out / name).exists():
+            shutil.copy2(src / name, out / name)
+
+    tc = out / "tokenizer_config.json"
+    if tc.exists():
+        cfg = json.loads(tc.read_text(encoding="utf-8"))
+        for k in _TRANSIENT_TOKENIZER_KEYS:
+            cfg.pop(k, None)
+        cfg["model_max_length"] = MODEL_MAX_LENGTH
+        tc.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    # Strip HF trust_remote_code files the convert step may have copied through;
+    # the MLX repo loads via model_file and shouldn't ship the PyTorch reference.
+    for name in _STRIP_FROM_OUTPUT:
+        p = out / name
+        if p.exists():
+            p.unlink()
+
+
+def verify(out_dir: str | Path) -> None:
+    """Assert the converted repo's tokenizer round-trips the EOS token.
+
+    A conversion that silently loses ``<|im_end|>`` produces a model that never
+    stops generating — cheap to assert, expensive to debug.
+    """
+    from transformers import AutoTokenizer  # type: ignore
+
+    out = Path(out_dir)
+    tok = AutoTokenizer.from_pretrained(str(out), trust_remote_code=False)
+    got = tok.convert_tokens_to_ids(EOS_TOKEN)
+    assert got == EOS_TOKEN_ID, (
+        f"EOS token lost in conversion: {EOS_TOKEN!r} -> {got} (expected {EOS_TOKEN_ID})"
+    )
+    # ``eos_token_id`` may be stored as a string in some tokenizer configs.
+    assert int(tok.eos_token_id) == EOS_TOKEN_ID, (
+        f"tokenizer.eos_token_id={tok.eos_token_id} != {EOS_TOKEN_ID}"
+    )
+    rendered = tok.apply_chat_template(
+        [{"role": "user", "content": "hi"}],
+        tokenize=False,
+        add_generation_prompt=True,
+    )
+    assert "<|im_start|>assistant" in rendered, "chat template missing assistant marker"
+    assert "<think>" in rendered, "chat template missing the reasoning (<think>) block"
+
+
+def to_mlx(
+    src_dir: str | Path,
+    out_dir: str | Path,
+    bits: int,
+    group_size: int = 64,
+    *,
+    skip_verify: bool = False,
+) -> Path:
+    """Convert ``src_dir`` to a quantized MLX repo at ``out_dir`` (``bits`` per weight).
+
+    Stages a non-mutating copy (P3.3), runs mlx-lm's convert, strips
+    ``auto_map`` from the output config, carries tokenizer files through and
+    verifies the round-trip (P3.4). The stage dir is always cleaned up.
+    """
+    src = Path(src_dir)
+    out = Path(out_dir)
+
+    stage = prepare_source(src)
+    try:
+        convert(
+            hf_path=str(stage),
+            mlx_path=str(out),
+            quantize=True,
+            q_group_size=group_size,
+            q_bits=bits,
+        )
+    finally:
+        shutil.rmtree(stage, ignore_errors=True)
+
+    # Strip auto_map from the output config: MLX loads via model_file, and
+    # leaving auto_map triggers an irrelevant trust_remote_code prompt on load.
+    cfg_path = out / "config.json"
+    cfg = json.loads(cfg_path.read_text(encoding="utf-8"))
+    cfg.pop("auto_map", None)
+    cfg_path.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    _finalize(src, out)
+
+    if not skip_verify:
+        try:
+            verify(out)
+            print(
+                f"PASS: tokenizer round-trips {EOS_TOKEN!r} -> {EOS_TOKEN_ID}; "
+                f"chat template intact."
+            )
+        except Exception as exc:  # pragma: no cover - surfaces as a hard error
+            print(f"VERIFY FAILED: {exc}")
+            raise
+
+    return out
+
+
+def main(argv: list[str] | None = None) -> None:  # pragma: no cover
+    import argparse
+
+    ap = argparse.ArgumentParser(
+        prog="nanbeige-mlx-convert",
+        description="Convert Nanbeige4.2-3B HF weights to a quantized MLX repo.",
+    )
+    ap.add_argument("--src", required=True, help="local Nanbeige HF repo (left untouched)")
+    ap.add_argument("--out", required=True, help="output MLX repo path")
+    ap.add_argument("--bits", type=int, required=True, help="quantization bits (4, 6, 8)")
+    ap.add_argument("--group-size", type=int, default=64)
+    ap.add_argument(
+        "--skip-verify",
+        action="store_true",
+        help="skip the post-convert tokenizer round-trip assertion",
+    )
+    a = ap.parse_args(argv)
+    to_mlx(a.src, a.out, a.bits, a.group_size, skip_verify=a.skip_verify)
+    print(a.out)
+
+
+if __name__ == "__main__":  # pragma: no cover
+    main()