Post-training a small Ling model with agentic RL on one DGX Spark

Ant Open Source showed Ling-3.0-tiny being post-trained, not just run, on a single Nvidia DGX Spark, using agentic reinforcement learning: the model plays a tic-tac-toe task, receives feedback from the environment and a reward signal, and adjusts its behavior from that feedback rather than from a fixed labeled dataset. Ant reports the average reward moved from about -0.5 to 0.4 over the course of training, and the model's response length dropped to roughly 850 tokens, both consistent with a model getting better at the task and communicating more concisely.

A companion Ant Ling post fills in specifics this thread omits: the training ran 400 steps using a method called GSPO, orchestrated through Ant's own AReno framework, with published tutorial and code. Reinforcement-learning post-training of agentic models is usually described as requiring a cluster of GPUs; demonstrating it end to end on one DGX Spark is the concrete claim here, though the tic-tac-toe task is a controlled, low-complexity environment rather than a real-world agent workload.