{
 "slug": "Tulu_3",
 "title": "Tulu 3",
 "type": "model",
 "short_desc": "Ai2's November 2024 family of openly post-trained Llama 3.1 derivatives, released with its full recipe and known for introducing reinforcement learning with verifiable rewards (RLVR).",
 "categories": [
  "2024 model releases",
  "Open-weight models",
  "Open-source AI",
  "Models"
 ],
 "infobox": {
  "Developer": "Allen Institute for AI (Ai2)",
  "Family": "Tülu (instruction-tuned model series)",
  "Announced": "November 2024",
  "Architecture": "Transformer decoder; fine-tuned from Llama 3.1 base models",
  "Parameters": "8B and 70B at launch; 405B added January 2025",
  "Post-training": "Supervised finetuning, length-normalized DPO, RLVR",
  "License / access": "Open weights (Llama 3.1 Community License); training data, code, and evaluation suite released",
  "Paper/report": "arXiv:2411.15124"
 },
 "infobox_links": {
  "Developer": [
   "Ai2"
  ],
  "Architecture": [
   "Transformer_(architecture)",
   "Llama_3.1"
  ],
  "Post-training": [
   "Direct_preference_optimization",
   "Reinforcement_learning_with_verifiable_rewards"
  ],
  "License / access": [
   "Llama_3.1"
  ]
 },
 "aliases": [
  "Tülu 3"
 ],
 "family": "",
 "org": "",
 "registry_status": "stub",
 "words": 380,
 "references": 3,
 "outbound": [
  "Ai2",
  "DeepSeek-V3",
  "Direct_preference_optimization",
  "GPT-4o",
  "Instruction_tuning",
  "Llama_3.1",
  "OLMo",
  "Qwen2.5",
  "Reinforcement_learning_from_human_feedback",
  "Reinforcement_learning_with_verifiable_rewards",
  "Transformer_(architecture)"
 ],
 "inbound": [
  "Ai2",
  "Flan-T5",
  "Molmo",
  "OLMo",
  "Reinforcement_learning_with_verifiable_rewards",
  "Stated_missions_of_AI_labs",
  "Supervised_fine-tuning"
 ],
 "url": "wiki/Tulu_3.html",
 "built": "2026-07-24"
}