Instructions to use cfli/A-base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cfli/A-base with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="cfli/A-base")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("cfli/A-base")
model = AutoModelForMultimodalLM.from_pretrained("cfli/A-base", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use cfli/A-base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cfli/A-base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfli/A-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cfli/A-base

SGLang

How to use cfli/A-base with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "cfli/A-base" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfli/A-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "cfli/A-base" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfli/A-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cfli/A-base with Docker Model Runner:
```
docker model run hf.co/cfli/A-base
```

cfli commited on 6 days ago

Commit

bc44530

verified ·

1 Parent(s): 887a6b2

Add files using upload-large-folder tool

Browse files

Files changed (2) hide show

README.md +12 -34
assets/arex-benchmark-results.svg +0 -0

README.md CHANGED Viewed

@@ -16,16 +16,13 @@ tags:
 ---
 <div align="center">
-  <img src="assets/arex-logo.png" width="52%" alt="AREX" />
-  <h2>AREX-Base</h2>
-  <p><strong>Towards a Recursively Self-Improving Agent for Deep Research</strong></p>
-[![Paper](https://img.shields.io/badge/Paper-AREX-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white)](#)
-[![GitHub](https://img.shields.io/badge/GitHub-AREX-24292F?style=for-the-badge&logo=github&logoColor=white)](https://545999961.github.io/AREX/web)
-[![AREX Demo](https://img.shields.io/badge/Live_Demo-AREX-0F766E?style=for-the-badge&logo=googlechrome&logoColor=white)](https://arex-research.com/)
-[![AREX-Turbo](https://img.shields.io/badge/Model-AREX--Turbo-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000)](https://huggingface.co/BAAI/AREX-Turbo)
 </div>
 ## Introduction
@@ -40,22 +37,22 @@ This repository contains **AREX-Base**, the larger model in the AREX family. It
 | :---: | :---: | :---: | :---: |
 | Mixture-of-Experts | 122B total / 10B activated | 262,144 tokens | Up to 300 research turns and 5 self-improvement rounds |
-## Highlights
 - **Recursive self-improvement:** evaluates provisional answers and chooses whether to accept, refine, or restart the research trajectory.
 - **Verification-guided research:** converts unresolved answer constraints into focused follow-up searches.
 - **Autonomous context updates:** retains verified findings, source identifiers, rejected candidates, open constraints, and the next research plan.
-- **Long-horizon tool use:** trained for multi-round search, browsing, evidence integration, and answer construction.
 - **Evidence-grounded answers:** maintains supporting evidence and answer-level confidence throughout the research process.
-## Model Family
 | Model | Architecture | Backbone | Context | Positioning |
 | --- | --- | --- | ---: | --- |
 | [AREX-Base](https://huggingface.co/BAAI/AREX-Base) | 122B total / 10B activated MoE | Qwen3.5-122B-A10B | 256K | Higher-capacity model with the strongest overall AREX results. |
 | [AREX-Turbo](https://huggingface.co/BAAI/AREX-Turbo) | 4B dense | Qwen3.5-4B | 256K | Compact model for lower-cost research and tool-use settings. |
-## How AREX Works
 AREX organizes deep research as two interacting loops supported by autonomous context management.
@@ -65,32 +62,17 @@ AREX organizes deep research as two interacting loops supported by autonomous co
 This design preserves useful work across iterations and reduces repeated exploration of paths that have already been rejected.
-## Training
-The AREX training pipeline combines:
-- verified synthetic research tasks that require multi-hop evidence acquisition and constraint checking
-- high-quality teacher trajectories filtered for coherent research behavior, valid tool use, evidence support, and answer confidence
-- agentic mid-training for browse-intensive tool use, evidence acquisition, expert reasoning, and hypothesis verification
-- long-horizon reinforcement learning with outcome rewards and step-level shaping at decision-critical research moments
-Step-level rewards target actions such as finding decisive evidence, rejecting an incorrect path, redirecting a search, and updating context after meaningful progress.
 ## Evaluation
 AREX is evaluated through a unified long-horizon search-agent interface with `search`, `visit`, `update_context`, and `finish` tools. HLE with tools additionally provides a Python tool. The reported setup allows up to 300 inner-loop turns and 5 outer-loop self-improvement operations.
-<div align="center">
-  <img src="assets/arex-benchmark-results.png" width="100%" alt="AREX benchmark results across six deep research benchmarks" />
-</div>
 AREX-Base scores **82.5** on BrowseComp, **85.4** on GAIA, **71.0** on xbench-2510, **89.9** on DeepSearchQA, **82.0** on WideSearch-en, and **52.4** on HLE with tools. It leads AREX-Turbo across all six reported benchmarks.
 <details>
 <summary><strong>Full benchmark table from the AREX paper</strong></summary>
-<br>
 **Frontier models**
 | Model | BrowseComp | GAIA | xbench-2510 | DeepSearchQA | WideSearch-en | HLE (tool) |
@@ -130,7 +112,3 @@ DeepSearchQA is reported with Item-F1. The other benchmarks are reported with ac
 ## Intended Use
 AREX-Base is intended for research-agent applications that require long-horizon information seeking, evidence aggregation, multi-constraint verification, and tool-augmented reasoning. Its larger capacity is suited to settings where research quality is prioritized over serving cost.
-## Limitations
-AREX depends on the availability and quality of its external search and browsing tools. It can still produce incorrect, incomplete, or weakly supported conclusions. Benchmark results also depend on the agent interface, tool implementation, inference configuration, and research budget described above. Outputs should be independently verified before high-stakes use.

 ---
 <div align="center">
+  <img src="assets/arex-logo.png" width="48%" alt="AREX"><br>
+  <strong>AREX-Base</strong><br>
+  <sub>Towards a Recursively Self-Improving Agent for Deep Research</sub><br>
+  <a href="#"><img src="https://img.shields.io/badge/Paper-AREX-B31B1B?style=flat-square&amp;logo=arxiv&amp;logoColor=white" alt="Paper"></a>
+  <a href="https://545999961.github.io/AREX/web"><img src="https://img.shields.io/badge/GitHub-AREX-24292F?style=flat-square&amp;logo=github&amp;logoColor=white" alt="GitHub"></a>
+  <a href="https://arex-research.com/"><img src="https://img.shields.io/badge/Live_Demo-AREX-0F766E?style=flat-square&amp;logo=googlechrome&amp;logoColor=white" alt="AREX Demo"></a>
+  <a href="https://huggingface.co/BAAI/AREX-Turbo"><img src="https://img.shields.io/badge/Model-AREX--Turbo-FFD21E?style=flat-square&amp;logo=huggingface&amp;logoColor=000" alt="AREX-Turbo"></a>
 </div>
 ## Introduction
 | :---: | :---: | :---: | :---: |
 | Mixture-of-Experts | 122B total / 10B activated | 262,144 tokens | Up to 300 research turns and 5 self-improvement rounds |
+**Key Features**
 - **Recursive self-improvement:** evaluates provisional answers and chooses whether to accept, refine, or restart the research trajectory.
 - **Verification-guided research:** converts unresolved answer constraints into focused follow-up searches.
 - **Autonomous context updates:** retains verified findings, source identifiers, rejected candidates, open constraints, and the next research plan.
+- **Long-horizon tool use:** supports multi-round search, browsing, evidence integration, and answer construction.
 - **Evidence-grounded answers:** maintains supporting evidence and answer-level confidence throughout the research process.
+**Model Family**
 | Model | Architecture | Backbone | Context | Positioning |
 | --- | --- | --- | ---: | --- |
 | [AREX-Base](https://huggingface.co/BAAI/AREX-Base) | 122B total / 10B activated MoE | Qwen3.5-122B-A10B | 256K | Higher-capacity model with the strongest overall AREX results. |
 | [AREX-Turbo](https://huggingface.co/BAAI/AREX-Turbo) | 4B dense | Qwen3.5-4B | 256K | Compact model for lower-cost research and tool-use settings. |
+## Method
 AREX organizes deep research as two interacting loops supported by autonomous context management.
 This design preserves useful work across iterations and reduces repeated exploration of paths that have already been rejected.
 ## Evaluation
 AREX is evaluated through a unified long-horizon search-agent interface with `search`, `visit`, `update_context`, and `finish` tools. HLE with tools additionally provides a Python tool. The reported setup allows up to 300 inner-loop turns and 5 outer-loop self-improvement operations.
+![AREX benchmark results across six deep research benchmarks](assets/arex-benchmark-results.svg)
 AREX-Base scores **82.5** on BrowseComp, **85.4** on GAIA, **71.0** on xbench-2510, **89.9** on DeepSearchQA, **82.0** on WideSearch-en, and **52.4** on HLE with tools. It leads AREX-Turbo across all six reported benchmarks.
 <details>
 <summary><strong>Full benchmark table from the AREX paper</strong></summary>
 **Frontier models**
 | Model | BrowseComp | GAIA | xbench-2510 | DeepSearchQA | WideSearch-en | HLE (tool) |
 ## Intended Use
 AREX-Base is intended for research-agent applications that require long-horizon information seeking, evidence aggregation, multi-constraint verification, and tool-augmented reasoning. Its larger capacity is suited to settings where research quality is prioritized over serving cost.

assets/arex-benchmark-results.svg ADDED Viewed