Bid Farewell to the Illusion of Local Models: Bailing Team Joins Forces with AReno to Establish a Local Intelligent Agent Reinforcement Learning Loop

The ability of a model to be deployed locally does not mean it can handle real-world business scenarios effortlessly. In specific business processes, AI not only needs basic conversation capabilities but must also accurately understand complex rules, adhere to security boundaries, and correctly call external tools to output results that meet specifications. To address this pain point, the Ant ASystem team, in collaboration with the open-source community, has maintained the local large model post-training toolkit AReno. Recently, AReno has partnered with the BaiLing large model to introduce a lightweight post-training practice path, successfully establishing an Agentic RL (reinforcement learning) loop in a single-machine environment.

To ensure the high reproducibility of the entire technical solution, this collaboration selected Tic-Tac-Toe as the minimal validation task, which has clear rules and direct feedback. The experiment was conducted on the DGX Spark hardware environment, performing post-training on the Ling-3.0-tiny model, which has a total of 7.9B parameters but only 1.3B activated parameters. At the beginning of training, the model could understand the basic rules, but still exhibited illegal actions during interaction. By converting the task rules into an accurate reward feedback mechanism and training for 400 steps using the GSPO algorithm, the model's average reward significantly increased, and the output length also converged. The retest results fully demonstrate that the model's stability and rationality in tool calling and action selection have made a qualitative leap.

The core value of this exploration lies in validating a transparent and reproducible task adaptation methodology: from accurately identifying errors, defining verifiable feedback, to completing local training and retesting in the same environment. This model is not only applicable to chess experiments but can also be smoothly migrated to diverse real-world production scenarios such as tool call repair, structured field extraction, domain instruction following, and intelligent agents in business processes.

Currently, Ling-3.0-tiny provides multiple open-source versions including BF16, FP8, and INT4. Developers can obtain and use them through Hugging Face or the Moka Community. They can also access the AReno open-source repository to participate in subsequent code development and technical discussions.