rao-studios/Qwen3-Coder-30B-A3B-Instruct-mlx-4bit

Rao Studios' mirror of mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit at commit 6e302ea604ad9ab206367e2c501d1571023e7b6d. 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/Qwen3-Coder-30B-A3B-Instruct at b2cff646eb4bb1d68355c01b18ae02e7cf42d120) 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.

Lineage

• Qwen/Qwen3-Coder-30B-A3B-Instruct, by the Qwen team (Apache-2.0).
• mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit: that model converted to MLX at 4 bits — the files here.

Shape (from config.json)

Architecture	qwen3_moe, 48 layers
Quantization	4-bit, group size 64
Weights	17.2 GB of safetensors, 1,351 tensors
Context	262,144 tokens
Tools	Qwen3-Coder XML: the chat template lists every tool as XML, and the model answers with a <tool_call> block whose parameters are plain text between tags

A call, as the model writes it:

<tool_call>
<function=edit_file>
<parameter=path>
Sources/App/main.swift
</parameter>
<parameter=old_string>
print("hi")
</parameter>
<parameter=new_string>
print("hello")
</parameter>
</function>
</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 6e302ea604ad9ab206367e2c501d1571023e7b6d.
• No auto_map (remote code) in the config or the tokenizer config; a chat template is present.
• qwen3coder_tool_parser.py was dropped: Python is never mirrored (Craft reads tool calls itself).

File	Bytes	SHA-256
added_tokens.json	707	c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680
chat_template.jinja	6,722	672e747c77e990320152343b0a4951222e40de5645297905d89afba05586d827
config.json	10,897	8200545374776769bfd5ae4a30533990c96ca1cad625a9500194f46a8549fa13
generation_config.json	180	c99723ab3ba28630d26ae23def77603b540a46924895af5cf234740d3b27b51d
merges.txt	1,671,853	8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5
model-00001-of-00004.safetensors	5,321,473,414	25c6d0a62e02224605774ce2e9dc9a8c95910cde33557b938e55a368da1b5212
model-00002-of-00004.safetensors	5,366,644,780	7007ce495b063155207c3e8b4038771d0d751655274cb6b76c42b8d9c0624412
model-00003-of-00004.safetensors	5,276,887,419	a14f5bfdd0cd72ee6e71e7f0bc72f6fcde4d78a2da7ee0665f8cf071ac7d2675
model-00004-of-00004.safetensors	1,216,066,381	0ea3484a221d674b60c604ea20cb5e5418a2cfd9e61df72c0d7fb5b1622ff5b4
model.safetensors.index.json	119,598	c100efb419233c6be585014ebd9f5a8c2a4fe9de63be6271172f0f70f1c315ea
special_tokens_map.json	613	76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
tokenizer.json	11,422,654	aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
tokenizer_config.json	5,405	05a2d456eb807c8368fe6b60101b1d1160d6889071a751ed08576b96e77e01cd
vocab.json	2,776,833	ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910

Use with mlx

pip install mlx-lm
mlx_lm.generate --model rao-studios/Qwen3-Coder-30B-A3B-Instruct-mlx-4bit --prompt "Write a binary search in Swift."