rao-studios/Qwen2.5-Coder-14B-Instruct-mlx-4bit

Rao Studios' mirror of mlx-community/Qwen2.5-Coder-14B-Instruct-4bit at commit 29efdbab55a161237ab1e432a3abaf6c7ae2b477. The weights, tokenizer, chat template and configuration are byte-for-byte that repository's; nothing was re-quantized or edited. This card and LICENSE (the Apache License 2.0, verbatim from Qwen/Qwen2.5-Coder-14B-Instruct at aedcc2d42b622764e023cf882b6652e646b95671) were added.

It exists so every install of Craft fetches the same bytes from a repository Rao controls: Craft's onboarding recommends it by this Mac's memory, and pins this repository at one commit.

Craft

Craft is a coding app that runs on your Mac: the model and the index of your project stay on the machine. Craft offers this model for a Mac with 24 GB: a dense model, steady on small changes, with a 32k context.

Craft builds a whole iPhone app from an empty folder in one turn. The model plans, then calls fifteen agents from Craft's store in order: the project, the look, the mark, the typed client, the mock world, each screen's kit, the navigation, the tests and the screenshots. Then it builds, tests and shoots. Measured twice, the app built, its tests passed, and all 62 of the original app's screens matched it, light and dark: the quality Claude Opus reached on the same app.

Those runs used Qwopus 3.5 122B on an M4 Max with 128 GB (2026-11-16), not this model. In Craft, Settings โ€บ Guide walks through the run, says which model and Mac it needs, and sets Craft up for it. More at craft.rao.nyc.

Lineage

Shape (from config.json)

Architecture qwen2, 48 layers
Quantization 4-bit, group size 64
Weights 8.3 GB of safetensors, 1,255 tensors
Context 32,768 tokens
Tools Hermes JSON: the chat template lists every tool's JSON schema inside <tools>, and the model answers with one JSON object per call inside <tool_call> tags

A call, as the model writes it:

<tool_call>
{"name": "edit_file", "arguments": {"path": "Sources/App/main.swift", "old_string": "print(\"hi\")", "new_string": "print(\"hello\")"}}
</tool_call>

Safety pass

Run by Craft's scripts/mirror-model.py on 2026-10-26, before upload:

  • Every file is one an MLX model needs; no pickle format and no native code was in the source.
  • Every weight file is safetensors, its header read by hand: known dtypes, tensors that tile the file exactly, the same tensor names as the index.
  • Every file matches the SHA-256 the Hub recorded for it at 29efdbab55a161237ab1e432a3abaf6c7ae2b477.
  • No auto_map (remote code) in the config or the tokenizer config; a chat template is present.
File Bytes SHA-256
added_tokens.json 605 58b54bbe36fc752f79a24a271ef66a0a0830054b4dfad94bde757d851968060b
config.json 867 7a2e85de5e0e115e6aaea7cb71091c5833aa29786fc748ef1d7e607e3a3a18f2
merges.txt 1,671,853 8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5
model-00001-of-00002.safetensors 5,353,840,273 784b0c89d3e006f49cb8547aba1f2f4e99b365ab4380cf202655ca04c87d7ac3
model-00002-of-00002.safetensors 2,955,653,960 a1b7516f1926070177601698661494dfc68c2de1d98e7c43ba910a0d1d7dfbbe
model.safetensors.index.json 107,336 635d985ebf00b6a25daba71b7eba4d5ec8be7902c15104a79312fe947f44b121
special_tokens_map.json 613 76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
tokenizer.json 7,031,673 a8506e7111b80c6d8635951a02eab0f4e1a8e4e5772da83846579e97b16f61bf
tokenizer_config.json 7,306 7e88129d9769a0b14b1587a7d5e829fe93ac0e1511636471fdfc0811951418e6
vocab.json 2,776,833 ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910

Use with mlx

pip install mlx-lm
mlx_lm.generate --model rao-studios/Qwen2.5-Coder-14B-Instruct-mlx-4bit --prompt "Write a binary search in Swift."
Downloads last month
417
Safetensors
Model size
15B params
Tensor type
U32
ยท
F16
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for rao-studios/Qwen2.5-Coder-14B-Instruct-mlx-4bit