mlx-community/AREX-2-4bit

AREX-2 for Mac. This is BAAI/AREX-2 converted to run on Apple Silicon Macs. AREX-2 is an open model for coding, reasoning and agent work, and it can read images as well as text. This is the 4-bit version, converted to MLX format with mlx-vlm 0.7.4.

• Original model: BAAI/AREX-2, built on Qwen3.8-27B
• Size: 27 billion parameters, 16 GB download
• Takes in: text and images. Gives back: text
• Thinks before answering, and supports tool calling
• License: Apache 2.0

Based on my testing, use this size only if the 5-bit does not fit your Mac. It needs about twice the room to think, so it is the slowest to finish an answer. The numbers and the settings it needs are below.

Speed and quality

Size	Writing speed	With a speed helper	Match to the original	Hard problems solved, first try	Within 3 tries
4-bit	14 tokens/s	27 tokens/s	91.4%	3 of 12	9 of 12
5-bit	12 tokens/s	20 tokens/s	95.9%	8 of 12	11 of 12
6-bit	10 tokens/s	19 tokens/s	97.6%	9 of 12	10 of 12
8-bit	8 tokens/s	19 tokens/s	99.1%	7 of 12	12 of 12

• Measured on a Mac mini M4 Pro with 64 GB. A token is about three-quarters of a word. Reading a prompt runs at about 90 tokens per second at every size.
• Loading took 18 to 33 seconds depending on size (18 for this one), from an external SSD.
• Speed helper: a small draft model that guesses ahead. See "Speed with a draft model" below.
• Match to the original: how often the build picks the same next word-piece as the full-size model.
• Hard problems: 12 Python problems, with up to 3 tries each and a 4,000-token limit. One run per size, so a gap of one or two is noise. With an 8,000-token limit the 4-bit solved 8 of 12 hard problems on the first try (the 5-bit: 11 of 12 at the same limit); with 4,000 it solved 3.

Which size should I download?

Your Mac's memory	Get this one	Download	Memory it used	In short
24 GB	4-bit	16 GB	16 to 24 GB	Smallest. Needs a response limit of 8,000 tokens or more, so it is the slowest to finish an answer.
32 GB	5-bit	19 GB	20 to 28 GB	Scored about the same as the 6-bit and 8-bit in the tests, in the least memory.
48 GB	6-bit	23 GB	23 to 31 GB	Slightly closer to the original, a little slower.
64 GB or more	8-bit	30 GB	30 to 38 GB	Closest to the original, slowest.

• Memory it used was measured on a 64 GB Mac, from a short chat up to a very long prompt (32,000 tokens).
• The Mac sizes in the first column are estimates from that memory use. Nothing was run on a smaller Mac.

How to run it

Easiest: LM Studio, no coding

1. Install LM Studio and open it.
2. Open its model search and paste https://huggingface.co/mlx-community/AREX-2-4bit.
3. Click Download. It is 16 GB.
4. Start a new chat, choose AREX-2 in the model picker, and wait for it to load.
5. Type your question. To ask about a picture, attach it to your message.

Command line

Install once:

pip install -U mlx-vlm

Ask a question. The first run downloads the model (16 GB):

python -m mlx_vlm.generate \
  --model mlx-community/AREX-2-4bit \
  --max-tokens 8000 --temperature 1.0 --top-p 0.95 --top-k 20 \
  --prompt "Write a Python function that merges overlapping intervals."

Ask about a picture by adding --image:

python -m mlx_vlm.generate \
  --model mlx-community/AREX-2-4bit \
  --max-tokens 8000 --temperature 1.0 --top-p 0.95 --top-k 20 \
  --prompt "What is in this picture?" --image photo.png

From other apps and agent harnesses

Start LM Studio's local server (the Developer tab, or lms server start). Any app that speaks the OpenAI API can then use http://localhost:1234/v1 with this model, including tool calls and images.

Which apps it runs in

App	Status
LM Studio 0.4.8	Tested: text, images and tool calls
mlx-vlm 0.7.4 (command line and Python)	Tested: text, images and tool calls
mlx-lm 0.32	Tested: text only
mlx-dspark 0.20.1	Tested: text, with a speed helper
MiniMax Code 0.5.9 (agent harness)	Tested on the 8-bit: it fixed three small failing projects using its own file and shell tools, in 6 of 6 trials. Served through mlx-dspark with thinking off.
Other agent harnesses and coding tools	Should work. Tool calling works, and LM Studio's local server returns standard tool calls, which is what most harnesses use. Not tested.
Jan, Msty and other apps with an MLX engine	Not tested. They load MLX models, so they may work. Image input depends on the app.
Ollama, llama.cpp and other GGUF apps	These need a different format. Use a GGUF build such as bartowski/BAAI_AREX-2-GGUF.

Two settings that matter

• Leave the sampling at the model's defaults: temperature 1.0, top-p 0.95, top-k 20. Do not set temperature to 0. At temperature 0 this size got stuck repeating </think> on 4 of 18 tasks. (If you turn thinking off, Qwen's guidance for the base model is temperature 0.7, top-p 0.8, top-k 20; that was not tested here.)
• Give it room to answer. The model thinks before it replies, and a hard problem can take 1,000 to 4,000 tokens of thinking, often more on this size. If answers get cut off, raise the response limit (max tokens) to 8,000 or more.

What the tests showed

• The 5-bit, 6-bit and 8-bit are too close to tell apart. The 4-bit is clearly weaker on hard problems.
• AREX-2 was more efficient than plain Qwen3.8-27B: about 46% fewer tokens and about half the time on the same tests.
• It can run more than twice as fast with a speed helper. The DFlash2 helper for Qwen3.8-27B (incoai/Qwen3.8-27B-DFlash2) works with AREX-2 as it is. In mlx-dspark it took the 8-bit from 8 to 19 tokens per second.
• Every size passed a four-step tool task, a 32,000-token recall test, 11 image questions, and loading in LM Studio.

The numbers are below. The scripts and raw results are in the tests folder, and on GitHub at MuscleOtter/arex-2-mlx-tests.

AREX-2 compared with plain Qwen3.8-27B

	AREX-2	Qwen3.8-27B
Tokens used, all tests	113,795	212,514
Time, all tests	249 min	451 min
Easy tasks passed	35 of 36	30 of 36
Hard problems, first try	15 of 24	11 of 24
Hard problems, within 3 tries	23 of 24	20 of 24

• Same conditions for both: 8-bit MLX, the same prompts, settings and 4,000-token limit, two runs of each test.
• The gap is mostly length. Every failed attempt by plain Qwen3.8-27B ran into the 4,000-token limit (31 of 31), against 9 of 13 for AREX-2. With a higher limit it might solve as many, more slowly.
• Both ran with thinking on, the default for both models.
• With thinking off, in a real harness, they were level. In MiniMax Code both passed 6 of 6 fix-the-tests trials, in 17 and 18 minutes. The efficiency gap above comes from thinking.
• This is a small test on one Mac, not a benchmark. It does not measure the long multi-round tasks AREX-2 was trained for.

Speed with a draft model

A draft model is a small helper that guesses ahead so the main model can write faster. It changes speed, not quality. The helper made for plain Qwen3.8-27B, incoai/Qwen3.8-27B-DFlash2, works with AREX-2 unchanged. Measured with mlx-dspark 0.20.1, six prompts each:

Size	Tokens/s without helper	Tokens/s with helper
4-bit	14	27
5-bit	12	20
6-bit	10	19
8-bit	8	19

The test problems, by name

Easy tasks (18): merge intervals, integer to Roman numeral, longest palindromic substring, reverse Polish notation, spiral matrix order, minimum window substring, decode nested string, next permutation, top-k frequent words, simplify Unix path, expression calculator, count islands, edit distance, largest rectangle in a histogram, trapping rain water, word break, course schedule order, longest increasing subsequence.

Hard problems (12): regular expression matching, text justification, strong password checker, valid number, number to English words, skyline, shortest palindrome (200,000 characters), count smaller numbers after self (100,000 elements), sliding window median (100,000 elements), burst balloons, trapping rain water in 2D, max points on a line.

Each is one Python function checked by hidden tests. These are well-known problems, so both models have probably seen them in training.

What was and was not checked

Every size passed all of these:

• A plain answer, with thinking on and off.
• A four-step task using tools, on 3 of 3 runs.
• Finding three facts hidden in a 32,000-token prompt.
• 11 questions about images: an invoice, a bar chart, counting shapes, two images at once, and fine print on a large image.
• Loading and chatting in LM Studio 0.4.8, with text, tool calls and images (checked on the 5-bit, including a copy downloaded back from this page).

The image questions were run at temperature 0. At the recommended settings the model read an invoice total correctly in 6 of 7 tries; once it misread a digit and then corrected itself.

Not checked: video, prompts longer than 32,000 tokens, Macs with less than 64 GB, or BAAI's own benchmarks.

For developers, and how it was made

• Tool calls come back as <tool_call><function=name><parameter=key>value</parameter></function></tool_call>, not JSON. LM Studio converts them to standard tool calls for you.
• Text only: the model also loads in mlx-lm.
• Conversion: mlx-vlm 0.7.4, plain quantization, 4.70 bits per weight, original chat template unchanged.

python -m mlx_vlm convert --hf-path BAAI/AREX-2 --mlx-path AREX-2-4bit -q --q-bits 4

Common questions

• Can AREX-2 run on a Mac? Yes. These MLX builds run on Apple Silicon Macs. They were tested on an M4 Pro with 64 GB.
• How much memory does AREX-2 need on a Mac? On the test Mac the 5-bit used about 20 to 28 GB, the 6-bit 23 to 31 GB and the 8-bit 30 to 38 GB. That suggests a Mac with 32 GB or more, though nothing was run on a smaller one.
• Which AREX-2 MLX size is best? The tests here could not separate the 5-bit, 6-bit and 8-bit. The 5-bit needs the least memory of the three; the 8-bit is closest to the original.
• Does AREX-2 work in LM Studio? Yes, including images and tool calls.
• Does it work in Ollama? Not this build. Ollama uses the GGUF format, so use a GGUF version of AREX-2 there. Other MLX apps such as Jan and Msty may work but were not tested.
• Can it read images? Yes. Image input was kept in the conversion.
• Is AREX-2 better than Qwen3.8-27B? In a small local test it solved a few more problems and used about 46% fewer tokens. That is not a benchmark.

Interested in a further distilled 4-bit?

The 4-bit here is a plain conversion, and it is the weakest size. A further distilled 4-bit might do better: the small build is trained to copy the 8-bit's answers, which can win back some of the quality lost at this size. It would stay the same size and keep image input. It takes a few days of computer time and may not close the gap, so I will only make it if people want it. If you are interested, leave a comment in the Community tab and tell me what Mac you have.

Questions and license

Found a problem or have a question? Open a discussion on the Community tab of this page.

Apache 2.0, the same as the original. Credit for the model goes to BAAI (paper).