Tulu 3 70B

Positioning

Tulu 3 70B is Allen Institute for AI's (AI2) 70B-parameter instruction-tuned model — a fully open pipeline from base model through instruction tuning, RLHF, and evaluation. Built on the Llama 3.1 70B base with AI2's open instruction-tuning methodology, Tulu 3 represents the "open science" approach to instruction-following: every component of the training pipeline (data, code, evaluation) is publicly released. For operators who value transparency, reproducibility, and fully open training pipelines, Tulu 3 is the reference 70B instruction-tuned model.

Strengths

• Fully open training pipeline. AI2 releases the training data, instruction-tuning recipes, RLHF methodology, and evaluation framework. Every component from base model to final tuned model is documented and reproducible. Unique among 70B-class instruction-tuned models.
• AI2's instruction-tuning research quality. The Tulu lineage (Tulu 1 → 2 → 3) has produced published research on instruction-tuning methodology, with Tulu 3 incorporating lessons from multiple generations of open instruction-tuning research.
• Licensing through the open pipeline. While the base Llama 3.1 model carries the Llama Community License, AI2's training artifacts (data, code, evaluation) are openly licensed. Operators can reproduce the training pipeline with different base models if Llama's license is a concern.
• Competitive instruction-following benchmarks. Tulu 3 70B's instruction-following performance is competitive with other 70B instruction-tuned models on AlpacaEval, MT-Bench, and general instruction-following benchmarks.
• Strong coding and reasoning from the Llama 3.1 base. The underlying 70B model provides competitive reasoning, coding, and general knowledge capability from Meta's training infrastructure.
• 70B scale fits consumer/workstation hardware. Q4 at ~42 GB fits Mac Studio M3 Ultra or dual RTX 4090.

Limitations

• Inherits Llama 3.1 Community License restrictions. The base model is Llama 3.1 70B, carrying the 700M MAU clause and Llama use restrictions. The open training pipeline is open, but the model weights carry Llama's license.
• Instruction-tuning quality is competitive but not class-leading. Dedicated instruction-following specialists like WizardLM-2 8x22B or API-hosted models (GPT-4, Claude) may outperform Tulu 3 on complex instruction-following tasks.
• Base model is Llama 3.1, not 3.3 or 4. The underlying base model is from the Llama 3.1 generation — knowledge cutoff is earlier and some Llama 3.3/4 architectural improvements are absent. The model's strength is the open training pipeline, not the base model generation.
• No vision or multimodal capability. Text-only. For instruction-tuned VL models, consider Llama 3.2 90B Vision Instruct.
• AI2 research focus vs production optimization. Tulu 3 is primarily a research output — the training pipeline is optimized for transparency and reproducibility, not for production inference throughput or latency. Production deployment may require additional optimization.

Real-world performance

• vs Llama 3.3 70B: Tulu 3 is instruction-tuned on Llama 3.1; Llama 3.3 is the next-generation base. Llama 3.3 70B likely has stronger base capability; Tulu 3 has AI2's open instruction-tuning pipeline advantage. Pick by base model generation vs open pipeline preference.
• vs Llama 3.1 Nemotron 70B Instruct: Both fine-tunes of Llama 3.1 70B. Nemotron 70B has NVIDIA's distillation and optimization pipeline; Tulu 3 has AI2's fully open training pipeline. Pick by optimization depth (Nemotron) vs openness (Tulu 3).
• vs WizardLM-2 8x22B: Different architectures (MoE vs dense) and different training approaches. WizardLM-2 likely has stronger instruction-following on complex tasks; Tulu 3 has the open pipeline transparency advantage.

Should you run this locally?

Yes if you value fully open, reproducible training pipelines over maximum benchmark performance, can accept the Llama 3.1 base license, and have workstation-class hardware. Tulu 3's "open science" approach is unique and valuable for operators who want to understand and potentially reproduce their model's training.

No if you need maximum instruction-following quality at any cost (pick WizardLM-2 8x22B or API alternatives), the Llama license is incompatible, or you prefer a more recent base model generation (pick Llama 3.3 70B).

How it compares

• vs Llama 3.3 70B: Tulu 3 is open-pipeline instruction-tuned Llama 3.1; Llama 3.3 is newer base generation.
• vs Llama 3.1 Nemotron 70B Instruct: Both Llama 3.1 fine-tunes. Nemotron for NVIDIA optimization; Tulu for open pipeline.
• vs WizardLM-2 8x22B: Different architectures and training philosophies. Different strengths.
• vs Qwen 2.5 72B Instruct: Qwen 2.5 72B uses the Qwen license; Tulu is Llama-based with open pipeline.

Run this yourself

• Datacenter: Single A100 80GB at AWQ-INT4 with vLLM.
• Workstation: Mac Studio M3 Ultra (96 GB or 256 GB) at Q4. The 43 GB file fits either.
• Consumer: Single RTX 5090 (32 GB) holds only part of the 43 GB Q4_K_M file, so it needs partial offload to system RAM.
• Vendor: huggingface.co/allenai/Llama-3.1-Tulu-3-70B