OLMo 2/3 + Tülu 3 (Ai2)
The fully-open comparison · Deep-Dive FTDD-02 · Course 3
45 minutes · The open base (OLMo) + the open post-training recipe (Tülu 3)
What "open" means when a research lab commits to releasing everything — and why it's the counterpoint to MiniCPM's product orientation.
Deep-Dives
Ai2 — the canonical fully-open lab
Two releases
- OLMo — the open BASE
- Tülu 3 — the open POST-TRAINING recipe
License
Apache-2.0 — weights, data, and code. The strict end of the FT02 open spectrum.
OLMo 2 (arXiv:2501.00656) established the fully-open standard. Tülu 3 (arXiv:2411.15124) released the full post-training pipeline. Together they let you rebuild and prove the whole model.
OLMo 3 — three variants, one base
Released November 2025 · 7B + 32B · the steer axis made concrete.
VARIANT 1
base
Pretrained weights. Layer 1 only. What you'd fine-tune from.
VARIANT 2
instruct
+ SFT + DPO. Layer 3 (format + preference). The general chat model.
VARIANT 3
think
+ RLVR. Layer 3 (+ reasoning). The reasoning-tuned variant.
Pedagogical gift: the same base, three steers, directly comparable. Load all three and see what each steering stage changed.
The Tülu 3 post-training pipeline
Three stages, each a Layer-3 steer (FT00 stack). Each stage's data, code, and config is released.
| Stage | Technique | Steers | Module |
|---|
| 1 | SFT | Format, instruction-following | FT12 |
| 2 | DPO | Preference alignment | FT13 |
| 3 | RLVR | Reasoning (verifiable rewards) | FT14 |
RLVR (Reinforcement Learning on Verifiable Rewards) is the technique behind the "reasoning model" wave — and Tülu 3 is the open implementation. The recipe lives in allenai/open-instruct.
Fully-open — what you actually get
Reproducibility
Rebuild from a pinned commit — prove no silent drift (FT02).
Ablation
Remove one stage (e.g., skip RLVR), measure the effect. Research advances on this.
Supply-chain trust
Audit end-to-end, rule out hidden training-time exfiltration (FT22).
OpenBMB vs Ai2 — two open philosophies
| Property | OpenBMB / MiniCPM | Ai2 / OLMo + Tülu |
|---|
| Orientation | Product / edge | Research |
| Sizes | 1B–4B + multimodal | 7B–405B, text-focused |
| Variants | Modality axis (text/vision/omni) | Steer axis (base/instruct/think) |
| Default use | Ship small model on edge hardware | Reproduce, audit, research the stack |
MiniCPM is what you SHIP. OLMo/Tülu is what you STUDY. Both open-data/open-recipe; different centers of gravity.
Anti-patterns
Treating "fully-open" as "production-ready." OLMo/Tülu are research artifacts first — fully open and auditable, but not always the most capable at their size. Choose them for openness and reproducibility, not leaderboard rank. Capability is FT03.
Assuming OLMo 3 "think" = GPT-class reasoning. It's a 7B/32B RLVR-tuned model — it demonstrates the technique openly, but won't match a closed frontier reasoner. Use it to UNDERSTAND reasoning fine-tuning (FT14), not to deploy a frontier reasoner.
Skipping the open-instruct repo. Tülu 3's value is the recipe, and the recipe lives in code. Reading the paper without the repo gives you the description, not the executable pipeline.
What you can now do
- Distinguish OLMo (open base) from Tülu 3 (open post-training recipe) and explain why both are needed for reproducibility.
- Map the OLMo 3 base/instruct/think variants to the steering stages (none, SFT+DPO, +RLVR).
- Trace the Tülu 3 pipeline (SFT → DPO → RLVR) onto the FT00 stack and cite both arXiv papers.
- Defend why fully-open matters for reproducibility, ablation, and supply-chain trust.
- Contrast OpenBMB (product/edge) and Ai2 (research/audit) philosophies.
The lab: compare OLMo 3's base, instruct, and think variants on a reasoning task — the three steering stages made visible on one base. No fine-tuning required.
Next: FTDD-03 — Unsloth