qwen3.5-4b-agentic-coder-v4-4bit-mlx

English | ๆ—ฅๆœฌ่ชž

Model Summary

This is a 4-bit quantized MLX conversion of mahiatlinux/qwen3.5-4b-agentic-coder-v4, a coding-focused fine-tune of Qwen3.5-4B tuned for agentic/coding assistant workloads.

This is an unofficial, community-produced conversion for running the model efficiently on Apple Silicon (M-series) Macs using the MLX framework. All credit for the fine-tune goes to its original author (mahiatlinux) -- thank you for releasing this model.

Usage

CLI

pip install -U mlx-lm
mlx_lm.generate --model masahiroid/qwen3.5-4b-agentic-coder-v4-4bit-mlx --prompt "Write a Python function that checks if a string is a palindrome." --max-tokens 512

Python API

from mlx_lm import load, generate

model, tokenizer = load("masahiroid/qwen3.5-4b-agentic-coder-v4-4bit-mlx")

prompt = "Write a Python function that checks if a string is a palindrome."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=512, verbose=True)
print(response)

Specs

Item Value
Base model mahiatlinux/qwen3.5-4b-agentic-coder-v4 (Qwen3.5-4B fine-tune)
Quantization 4-bit, group size 64
Framework MLX
Model size ~2.4 GB
Target hardware 8GB-class GPU, or Apple Silicon with 16GB+ unified memory

Intended use

This is a small (4B) model intended for lightweight local coding assistance -- autocomplete, simple function/snippet generation, boilerplate code -- rather than complex multi-file reasoning or architectural design work. For heavier tasks, a larger local model or a commercial service (e.g. GitHub Copilot) will generally give better results.

Notes

  • This conversion was produced with mlx_lm.convert --q-bits 4 --q-group-size 64.
  • This is a community conversion, not an official release from the original fine-tune author.

Security

Audited against its upstream with model-audit-lite: weight format, chat template, bundled code, and a machine-readable lineage (ML-BOM). Details, checksums and how to reproduce: SECURITY.md.


ใƒขใƒ‡ใƒซใฎๆฆ‚่ฆ

ๆœฌใƒขใƒ‡ใƒซใฏใ€Qwen3.5-4Bใ‚’ใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐใƒปใ‚จใƒผใ‚ธใ‚งใƒณใƒˆ็”จ้€”ๅ‘ใ‘ใซใƒ•ใ‚กใ‚คใƒณใƒใƒฅใƒผใƒณใ—ใŸ mahiatlinux/qwen3.5-4b-agentic-coder-v4 ใ‚’ใ€Apple Silicon็’ฐๅขƒ๏ผˆMLXใƒ•ใƒฌใƒผใƒ ใƒฏใƒผใ‚ฏ๏ผ‰ๅ‘ใ‘ใซ4-bit้‡ๅญๅŒ–ใ—ใŸ้žๅ…ฌๅผใฎใ‚ณใƒŸใƒฅใƒ‹ใƒ†ใ‚ฃๅค‰ๆ›็‰ˆใงใ™ใ€‚ ใƒ•ใ‚กใ‚คใƒณใƒใƒฅใƒผใƒณใฎๅˆถไฝœ่€…๏ผˆmahiatlinuxๆฐ๏ผ‰ใซๆ„Ÿ่ฌใ„ใŸใ—ใพใ™ใ€‚

ๅฎŸ่กŒๆ–นๆณ•

CLI

pip install -U mlx-lm
mlx_lm.generate --model masahiroid/qwen3.5-4b-agentic-coder-v4-4bit-mlx --prompt "Write a Python function that checks if a string is a palindrome." --max-tokens 512

Python API

from mlx_lm import load, generate

model, tokenizer = load("masahiroid/qwen3.5-4b-agentic-coder-v4-4bit-mlx")

prompt = "Write a Python function that checks if a string is a palindrome."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=512, verbose=True)
print(response)

ใ‚นใƒšใƒƒใ‚ฏไป•ๆง˜

Item Value
ใƒ™ใƒผใ‚นใƒขใƒ‡ใƒซ mahiatlinux/qwen3.5-4b-agentic-coder-v4๏ผˆQwen3.5-4Bใฎใƒ•ใ‚กใ‚คใƒณใƒใƒฅใƒผใƒณ๏ผ‰
้‡ๅญๅŒ–่จญๅฎš 4-bit, group size 64
ใƒ•ใƒฌใƒผใƒ ใƒฏใƒผใ‚ฏ MLX
ใƒขใƒ‡ใƒซใ‚ตใ‚คใ‚บ ็ด„2.4 GB
ๆƒณๅฎš็’ฐๅขƒ 8GB็ดšGPUใ€ใพใŸใฏApple Silicon 16GBไปฅไธŠใฎใƒฆใƒ‹ใƒ•ใ‚กใ‚คใƒ‰ใƒกใƒขใƒช

ๆƒณๅฎš็”จ้€”

ใ“ใฎใƒขใƒ‡ใƒซใฏ4Bใจใ„ใ†ๆฏ”่ผƒ็š„ๅฐใ•ใ„ใ‚ตใ‚คใ‚บใฎใŸใ‚ใ€่ปฝ้‡ใชใƒญใƒผใ‚ซใƒซใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐๆ”ฏๆด ๏ผˆ่‡ชๅ‹•่ฃœๅฎŒใƒป็ฐกๅ˜ใช้–ขๆ•ฐ/ใ‚นใƒ‹ใƒšใƒƒใƒˆ็”Ÿๆˆใƒปๅฎšๅž‹ใ‚ณใƒผใƒ‰ใฎ่จ˜่ฟฐ๏ผ‰ใ‚’ๆƒณๅฎšใ—ใŸ็”จ้€”ใซๅ‘ใ„ใฆใ„ใพใ™ใ€‚ ่ค‡ๆ•ฐใƒ•ใ‚กใ‚คใƒซใ‚’ใพใŸใ„ใ ่จญ่จˆๅˆคๆ–ญใ‚„่ค‡้›‘ใชใƒญใ‚ธใƒƒใ‚ฏใฎๅฎŸ่ฃ…ใซใฏใ€ใ‚ˆใ‚Šๅคงใใชใƒญใƒผใ‚ซใƒซใƒขใƒ‡ใƒซใ‚„ ๅ•†็”จใ‚ตใƒผใƒ“ใ‚น๏ผˆGitHub Copilotใชใฉ๏ผ‰ใฎๅˆฉ็”จใ‚’ๆŽจๅฅจใ—ใพใ™ใ€‚

ๅ‚™่€ƒ

  • ๆœฌๅค‰ๆ›ใฏ mlx_lm.convert --q-bits 4 --q-group-size 64 ใ‚’็”จใ„ใฆไฝœๆˆใ•ใ‚ŒใŸ้žๅ…ฌๅผใฎใ‚ณใƒŸใƒฅใƒ‹ใƒ†ใ‚ฃ็‰ˆใงใ™ใ€‚ใƒ•ใ‚กใ‚คใƒณใƒใƒฅใƒผใƒณใฎไฝœ่€…ใซใ‚ˆใ‚‹ๅ…ฌๅผใƒชใƒชใƒผใ‚นใงใฏใ‚ใ‚Šใพใ›ใ‚“ใ€‚

ใ‚ปใ‚ญใƒฅใƒชใƒ†ใ‚ฃใƒผ

model-audit-lite ใงๅค‰ๆ›ๅ…ƒใจ็ชใๅˆใ‚ใ›ใฆ็›ฃๆŸปๆธˆใฟใงใ™๏ผˆ้‡ใฟใฎๅฝขๅผใ€chat templateใ€ๅŒๆขฑใ‚ณใƒผใƒ‰ใ€ๆฉŸๆขฐๅฏ่ชญใช็ณป่ญœ๏ผML-BOM๏ผ‰ใ€‚่ฉณ็ดฐใƒปใƒใ‚งใƒƒใ‚ฏใ‚ตใƒ ใƒปๅ†็พๆ–นๆณ•ใฏ SECURITY.md ใ‚’ใ”่ฆงใใ ใ•ใ„ใ€‚

Downloads last month
2,031
Safetensors
Model size
4B params
Tensor type
U32
ยท
BF16
ยท
F32
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 masahiroid/qwen3.5-4b-agentic-coder-v4-4bit-mlx

Finetuned
Qwen/Qwen3.5-4B
Quantized
(4)
this model