OLMo 2/3 + Tülu 3 (Ai2)
The fully-open comparison: Allen Institute's OLMo (the open base) and Tülu 3 (the open SFT→DPO→RLVR post-training recipe). The research-oriented counterpoint to MiniCPM's product orientation.
OLMo and Tülu together cover the full open stack: the base AND the post-training recipe. OLMo (arXiv:2501.00656) is the open base — weights, data, code, checkpoints, eval. Tülu 3 (arXiv:2411.15124) is the open post-training pipeline — SFT, DPO, RLVR, each released as separable components. Both needed for full reproducibility.
OLMo 3's three variants (base/instruct/think) are a pedagogical gift. The same base with three different steers (none / SFT+DPO / +RLVR), directly comparable. A student can load all three and SEE what each steering stage changed — the FT00 steering stack made concrete, the clearest illustration of FT12/FT13/FT14.
Tülu 3's released pipeline enables ablation — the property that weights-only releases cannot provide. Because each stage (SFT, DPO, RLVR) is released as data + code + configs, a researcher can train Tülu 3 with one stage removed and measure the effect. This is how research advances, and it is impossible when the data and recipe are opaque.
MiniCPM is what you ship; OLMo/Tülu is what you study. Both open-data/open-recipe (Apache-2.0), but different centers of gravity: MiniCPM's variant axis is modality (product/edge); OLMo 3's variant axis is the steer (research/audit). MiniCPM for small edge models; OLMo/Tülu for reproducing, auditing, and understanding the full stack.