“A high-quality, reproducible code pretraining corpus comprising 960 Billion tokens across 607 programming languages.”
OpenCoder ships 1.5B and 8B models trained on 2.5 trillion tokens, 90% of it raw code. The point is the rest: the data pipeline, the pretraining set, the SFT data and intermediate checkpoints. Most “open” models hand over weights and hide the recipe. This one hands over the recipe, which is the only part anyone can check.