← Back to directory
Allen AI · Released 2025-12

Tulu 3

AI2's instruction-tuned Llama variant. Open recipe — every step of the post-training pipeline is documented and reproducible.

LlamaCommercial with caveatsgeneral
Params (max)
70B
Variants
70B / 8B
Context window
8K tokens
MMLU
75
HumanEval
65.6
GSM8K
87.1
Min VRAM (fp16, smallest variant)
16GB
Smallest Q4 GGUF
~4GB
Languages supported
4
Pros
  • ✓Open recipe
  • ✓Tuned for instruction following
  • ✓Backed by AI2
Cons
  • ×Inherits Llama license
  • ×Same context as base

Highlights

  • ●Fully reproducible post-training
  • ●Strong instruction following
  • ●Open data + code

Where to download

Hugging Face: allenai/Llama-3.1-Tulu-3-70B
Or via Ollama (ollama pull tulu-3) or LM Studio's in-app browser.
Homepage: https://allenai.org/tulu

Related reading

Fine-Tuning an Open Source LLM in 2026: LoRA vs QLoRA vs Full Fine-Tune
Should you LoRA, QLoRA, or full fine-tune your open LLM? Honest tradeoffs, GPU requirements, and a decision tree.