Tülu 3
The fully open post-training recipe that closed the gap with closed models.
About
Tülu 3 is AI2's model family built to prove that open post-training can match closed labs: the full four-stage recipe (data curation, SFT, DPO, and RLVR reinforcement learning on verifiable rewards) is published with code, data mixes, and evals. The 70B model posted strong numbers across standard benchmarks, and the 405B release was reported competitive with DeepSeek V3 and GPT-4o on key tasks. The Open Instruct codebase is Apache-2.0, though individual weight variants carry their base-model terms, so check the card for the size you want. It is a research artifact first, so expect to do your own serving and safety work.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Tülu 3 70B: MMLU 83.1, GSM8K 93.5, HumanEval pass@10 92.4 (AI2); 405B reported competitive with DeepSeek V3 and GPT-4o on key benchmarks (AI2, 2025).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Released 2024-11-21 by AI2; full recipe (SFT, DPO, RLVR) published with code and data; Open Instruct codebase Apache-2.0; weight variants carry base-model terms; 405B release reported competitive with DeepSeek V3 and GPT-4o (AI2, 2025); Research artifact; serving and safety work left to the user. License: Codebase Apache-2.0; check the per-variant HF card for weight licensing (Llama 3.1 lineage)..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Released 2024-11-21 by AI2; full recipe (SFT, DPO, RLVR) published with code and data
- · Open Instruct codebase Apache-2.0; weight variants carry base-model terms
- · 405B release reported competitive with DeepSeek V3 and GPT-4o (AI2, 2025)
- · Research artifact; serving and safety work left to the user