AREX-2

AREX-2

Improves its own solutions over several rounds of propose, measure, reflect and revise, using a 27B model from BAAI. For agent researchers and developers working on coding, machine-learning and deep-research tasks.

Made by BAAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AREX-2 is an agent model with 27B parameters, released by BAAI, the Beijing Academy of AI. It is built for long tasks where it improves an answer over many rounds. In each round it proposes a solution, measures it, reflects on the result and revises. It takes images and text and returns text.

Who it suits: It suits researchers and engineers who build agents that iterate on coding, machine-learning or research tasks with measurable feedback.

Industry | Who uses it | How they use it | Example
AI research | Agent researchers | Study self-improvement at test time | Compare results after more rounds on the same task
Software | ML engineers | Let a model refine code against scores and timings | An algorithm that gets faster after each round
Data science | Analysts | Iterate on machine-learning tasks with logs and errors | A training script fixed after reading its error logs
Research teams | Deep-research builders | Reuse the learned iteration habit for research tasks | A research agent that revises its answer across rounds

If your task has no score, log or other feedback to read, you can skip it.

Where to use it:

• Python with the transformers library
• Agent setups that run code or tools and return scores, logs and errors to the model
• Windows, macOS or Linux machines with enough GPU memory for 27B parameters

Use case | Who | What they get back
Algorithmic programming | Developers | A solution refined over several rounds
Machine-learning engineering | ML engineers | Code changes guided by scores and logs
Deep research | Researchers | Answers revised as the task budget grows

How to set it up: You need these first:

• Python and pip
• A GPU setup with room for about 54.7 GB of BF16 weights, or about 13.7 GB at a 4-bit estimate
• A way to run your task and send results back to the model

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load the model with transformers as the model card describes.

pip install -U huggingface_hub
hf download BAAI/AREX-2

How to use it: Give the model a task and a way to measure its work, then let it go through several rounds. Each round it reads what came back and decides what to change next.

Tool | What it does
Propose | Writes a candidate solution for the task
Measure | Runs the candidate and collects scores, logs, errors and timings
Reflect | Reads that feedback to decide what to change
Revise | Updates the solution and starts the next round

Example: You give it a slow routine and a timing script, and ask it to make the routine faster over several rounds. It would propose a version, read the timings, reflect on what held it back, and revise the code each round.

For newcomers and experts: Think of it as a model that checks its own work and tries again. A first small thing to try is a short coding task with a script that prints a score, and watch how the answer changes across rounds.

It is a dense multimodal model compatible with Qwen3.8, with a 262,144-token context. Training covered ML engineering and algorithm coding tasks whose results can be checked, along with the earlier AREX deep-research data. The authors say the learned habit carries over to deep research without new search data.

The model proposes a solution
Your setup measures the result
The model reflects on the feedback
It revises and starts the next round

Good to know: The card compares the model with closed-weight and open-weight models on several benchmarks, including Frontier-CS and MLE-Lite.

• Context: 262,144 tokens
• Size: Hugging Face lists 27.4B parameters stored as Safetensors
• Feedback: the method relies on scores, logs, errors and timings, so tasks with verifiable feedback suit it best
• Other versions: 16 quantized and 1 fine-tuned version made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the GPU
• Data it reaches: the task files, scores and logs you pass to it

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,028
• Likes: 208
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 18
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, agent, deep-research, reasoning, tool-use, long-context, self-improvement, conversational

Links

• Model: https://huggingface.co/BAAI/AREX-2
• Files: https://huggingface.co/BAAI/AREX-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BAAI/AREX-2

Then follow the usage in its model card. Nothing in this kit runs by itself.