Live artex-coder-7b is now on Hugging Face

concise · accurate · to the point

Less talk.
Better code.

Artex is a coding assistant that answers in code, not essays. Runs locally on your Mac — your code never leaves the machine.

artex-coder-7b Local · MLX

Demo answers are pre-written. Artex may make mistakes — review before you ship.

now live on hugging face

artex-coder-7b

0XARTEX/artex-coder-7b

Open weights. Fine-tuned from Qwen2.5-Coder-7B-Instruct to answer in code, not essays — in English or Bahasa Indonesia.

Text Generation
Transformers
Safetensors
Apache-2.0
EN · ID

7B: parameters
bf16: 15.2 GB weights
0.172: val loss (from 0.286)

View on Hugging Face Run it locally

quickstart.py

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "0XARTEX/artex-coder-7b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)

msgs = [{"role": "user", "content": "Debounce in TS"}]
ids = tok.apply_chat_template(
    msgs, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(ids, max_new_tokens=512)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

// why artex

Same question. A tenth of the words.

Most assistants pad every answer with preamble, caveats and recaps. Artex is tuned to skip all of it.

Typical assistant117 words

> How do I reverse a string in Python?

Great question! Reversing a string is a common task in Python, and there are actually several ways to accomplish it. In this answer, I'll walk you through the most popular approaches, explain how each one works, and discuss when you might want to use them.

The most Pythonic approach uses slicing. Slicing allows you to extract a portion of a sequence, and by providing a step of -1 you can traverse the string backwards…

Another approach is to use the built-in reversed() function combined with join(). This is slightly more verbose but some developers find it more readable…

In summary, both approaches work well. I hope this helps! Let me know if you have any other questions.

Artex8 words

> How do I reverse a string in Python?

s[::-1]

Slicing with step -1. Works on any sequence.

// features

chat

you › regex for a hex color
artex › /^#(?:[0-9a-f]{3}){1,2}$/i
        Matches #fff and #ffffff.

Answers, not essays.

Code first, one line of context, done. No "Great question!", no recap paragraph, no wall of disclaimers.

pytest

$ uv run pytest
collected 13 items

test_build_dataset.py ......
test_prepare_data.py  .......

13 passed

Trained on real, working code.

Fine-tuned from Qwen2.5-Coder-7B on curated, MIT-licensed instruction data. Loss only counts on answers, so it learns to respond — not to ramble.

zsh

$ mlx_lm.generate --model mlx_model \
    --prompt "retry fetch hook in React"
==========
export function useRetryFetch(url, tries = 3) {
  …
// generated on-device

Runs on your Mac. Stays on your Mac.

Convert the weights to 4-bit MLX with one command and Artex runs natively on Apple Silicon. No API key, no telemetry, no code uploaded anywhere.

chat

you › cara cek file ada di Node?
artex › import { existsSync } from "node:fs";
        existsSync("path") // true / false
        Versi async: fs.promises.access.

Replies in your language.

Ask in English, Bahasa Indonesia, or whatever you think in. Artex answers in kind — the code stays code.

PythonTypeScriptRustGoSQLSwiftBashJavaC++KotlinRubyPHP

// how it's built

Small model. Sharp edges.

A 7B coder, fine-tuned with LoRA in about 18 minutes on a single H100. Nothing exotic — just a tight pipeline.

01
Curate
4,000 Magicoder OSS-Instruct examples plus Artex identity chats (EN/ID).
02
Fine-tune
LoRA r16 · α32 on all linear layers, 2 epochs. Only answers count toward the loss.
03
Fuse
Adapters merged into the base weights — one standalone model.
04
Run
Published on Hugging Face. Load it and start asking.

7B: parameters
bf16: safetensors
15.2 GB: full weights
Apache-2.0: license

// get started

Three commands to your first answer.

Install MLX

pip install mlx-lm

Get 4-bit weights Live

mlx_lm.convert --hf-path 0XARTEX/artex-coder-7b -q

Ask

mlx_lm.chat --model mlx_model

// faq

Frequently asked.

What is Artex?

A coding assistant model fine-tuned from Qwen2.5-Coder-7B-Instruct to give short, accurate answers. It runs on your own Mac with MLX.

What hardware do I need?

The full bf16 weights are 15.2 GB — fine for a GPU or a high-memory Mac. Converted to 4-bit with MLX it shrinks to roughly 4 GB, so a 16 GB Apple Silicon Mac is comfortable.

Does my code get sent anywhere?

No. Inference happens entirely on your machine. There is no server to send it to.

Is it free?

Yes. The weights are on Hugging Face under Apache-2.0, the same license as the base model. Follow @Artex_llm for updates.

Has it been benchmarked?

Not yet. No HumanEval or MBPP run so far; expect quality close to the base model, with much shorter answers.

Can it replace a frontier model?

Not for everything. It's a 7B model — great for everyday snippets, fixes and explanations, weaker on large multi-file reasoning.

Stop reading. Start shipping.

Artex gives you the code and gets out of your way.

Get Artex Follow on X

Less talk.Better code.