Show HN: AIBuildAI Agent 2.0 – an AI agent that automatically builds AI models

AIBuildAI Agent 2.5 achieved #1 on OpenAI's MLE-Bench with a score of 70.7%, substantially outperforming the agents ranked 2nd through 5th. The AI agent, which automatically builds AI models, also ranked in the top 5.7% of 3,219 teams in Kaggle's TGS Salt Identification Challenge and in the top 6.6% of 4,370 teams in a heart disease prediction competition. Version 2.5 is available to Max users with features including a live build dashboard, run replay, cross-run memory, MCP research tools, and DeepSeek support.

🏆 1 on OpenAI MLE-Bench https://github.com/openai/mle-bench MLE-Bench https://github.com/openai/mle-bench 6/17/2026 The AIBuildAI Agent 2.5 version is made available to Max users https://www.aibuildai.io/ products : a live build dashboard, run replay, cross-run memory, MCP research tools, and DeepSeek support. Release notes and download https://github.com/aibuildai/AI-Build-AI/releases/tag/v2.5.1 . 5/26/2026 The AIBuildAI Agent 2.0 version is made available to Pro users https://www.aibuildai.io/ products . 5/1/2026 In the TGS Salt Identification Challenge https://www.kaggle.com/competitions/tgs-salt-identification-challenge hosted by Kaggle, the model automatically developed by our AIBuildAI Agent ranked in the top 5.7%. Among 3,219 teams composed of human experts, this performance reaches the level of top-tier human AI experts. Model and code developed by the Agent: tasks/tgs-salt-identification-challenge https://github.com/aibuildai/AI-Build-AI/tree/main/tasks/tgs-salt-identification-challenge . 4/27/2026 Excited to announce AIBuildAI Agent 2.0 It has once again achieved 1 on OpenAI's MLE-Bench, reaching a score of 70.7% and substantially outperforming the agents ranked 2nd through 5th. Compared to version 1.0, Agent 2.0 introduces several technical advancements, which we will detail in an upcoming technical report. The 2.0 version will be available to Pro users soon. 4/24/2026 In a heart disease prediction competition https://www.kaggle.com/competitions/playground-series-s6e2/overview hosted by Kaggle, the model automatically developed by our AIBuildAI Agent ranked in the top 6.6%. Among 4,370 teams composed of human experts, this performance reaches the level of top-tier human AI experts. Model and code developed by the agent: tasks/playground-series-s6e2 https://github.com/aibuildai/AI-Build-AI/tree/main/tasks/playground-series-s6e2 . demo.mp4 AIBuildAI is an AI agent that automatically builds AI models. Given a task, it runs an agent loop that analyzes the problem, designs models, writes code to implement them, trains them, tunes hyperparameters, evaluates model performance, and iteratively improves the models. By automating the model development workflow, AIBuildAI reduces much of the manual effort required to build AI models. On OpenAI MLE-Bench https://github.com/openai/mle-bench , AIBuildAI ranked 1, demonstrating strong performance on real-world AI model building tasks. AIBuildAI requires a Linux x86 64 machine Ubuntu 20.04 or newer . There are three versions. V2.5 Max is the current, most capable version. V2 Pro and V1 free remain available. | Version | Plan | To run it | |---|---|---| V2.5 Max — current | Max subscription | aibuildai login Max plan + a model API key Anthropic or DeepSeek | V2 Pro | Pro subscription | aibuildai login Pro plan + an Anthropic API key | V1 | free | an Anthropic API key no account | Subscriptions are managed at accounts.aibuildai.io https://accounts.aibuildai.io ; see the account page for the available plans. - Subscribe. Create an account at accounts.aibuildai.io/sign-up https://accounts.aibuildai.io/sign-up and switch to the Max plan. - Install. curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v2.5-latest/aibuildai-linux-x86 64.tar.gz | tar xz && ./aibuildai-linux-x86 64- /install.sh - Log in required before running : aibuildai login opens a browser to sign in aibuildai whoami should show an active Max plan - Set your model-provider API key. This is separate from the aibuildai subscription above: the subscription aibuildai login lets you run the product; this key pays for the AI model calls. aibuildai uses Anthropic Claude by default, so set your Anthropic API key: export AIBUILDAI API KEY=your-anthropic-api-key DeepSeek is also supported — set a deepseek- id in the config's llm.model and put your DeepSeek key in the same AIBUILDAI API KEY . - Run. V2.5 is driven by a YAML config: aibuildai config task.yaml writes a starter config with every field edit task.yaml: set run.task name, run.data root, run.instruction, run.playground root aibuildai run task.yaml Other commands: aibuildai memorize summarize past runs into memory , aibuildai replay <run-dir replay a finished run , aibuildai --help . - Subscribe to the Pro plan at accounts.aibuildai.io/sign-up https://accounts.aibuildai.io/sign-up . - Install. curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v2.0-latest/aibuildai-linux-x86 64.tar.gz | tar xz && ./aibuildai-linux-x86 64- /install.sh - Log in. aibuildai login aibuildai whoami should show an active Pro plan - Set credentials. export ANTHROPIC API KEY=your-api-key - Run. V2 is driven by command-line flags: aibuildai run --task-name <name --data-dir <path \ --playground-dir <path --instruction "$ cat task.md " --no-form Or run aibuildai with no flags to fill in the parameters in an interactive form. No account or subscription required. - Install. curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v1.0-latest/aibuildai-linux-x86 64.tar.gz | tar xz && ./aibuildai-linux-x86 64- /install.sh - Set your Anthropic API key. export ANTHROPIC API KEY=your-api-key - Run with command-line flags: aibuildai --task-name <name --data-dir <path \ --playground-dir <path --instruction "$ cat task.md " --no-form Or run aibuildai with no flags to fill in the parameters in an interactive form. Important: run the command directly in your terminal. Do not wrap it in a .sh / .bash script — running it through a script may cause the TUI Text User Interface to crash. After a run completes, the output directory usually looks like structure may slightly vary by task : ├── candidate 1/ candidate 2/ candidate 3/ Auto-generated training scripts and model checkpoints ├── checkpoint.pth Best model checkpoint ├── inference.py Standalone inference script for the final model ├── submission.csv Test predictions if test inputs are provided └── progress.pdf Visual progress report The main outputs of an AIBuildAI run are the model checkpoints and the script inference.py , which runs predictions with the final model on any data. When the task data folder includes unlabeled test inputs, AIBuildAI also writes a predicted-label file submission.csv . We provide ready-to-run task markdowns and datasets in the tasks/ folder of this repository. You can also write your own task description and point the run at your own dataset. git clone https://github.com/aibuildai/AI-Build-AI.git This project is licensed under the MIT License - see the LICENSE /aibuildai/AI-Build-AI/blob/main/LICENSE file for details. @article{zhang2026aibuildai, title={AIBuildAI: An AI Agent for Automatically Building AI Models}, author={Ruiyi Zhang and Peijia Qin and Qi Cao and Li Zhang and Pengtao Xie}, year={2026}, journal={arXiv}, url={https://arxiv.org/abs/2604.14455} } @article{zhang2026aibuildai2, title={AIBuildAI-2: A Knowledge-Enhanced Agent for Automatically Building AI Models}, author={Ruiyi Zhang and Peijia Qin and Qi Cao and Li Zhang and Pengtao Xie}, year={2026}, journal={arXiv}, url={https://arxiv.org/abs/2605.27873} }