whatmodel.sh / Llama 3.1 Tulu 3 70B DPO

Llama 3.1 Tulu 3 70B DPO

+ contribute

Llama 3.1 Tulu 3 70B DPO is an open-weights language model from Ai2, released November 2024. Its published size is 70.55B parameters and a 131k-token context window. Licence: llama3.1.

Catalogue identity notes

Comparison model from the Epoch AI internal-evaluation CSVs (Epoch release date 2024-11-21). Intermediate DPO checkpoint allenai/Llama-3.1-Tulu-3-70B-DPO; not the final Tulu 3 70B (RLVR) release (arena 'llama-3.1-tulu-3-70b', ollama tulu3:70b). 70,553,837,568 parameters; safetensors total 141.11 GB.

53.1±1.1#296 of 719 LLMsfrom 12 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of Llama 3.1 Tulu 3 70B DPO, out of 828 with a score, then more from Ai2.

#	model	score	where	RAM	released
353	GPT-4o mini (2024-07-18) OpenAI	53±0	api $0.26/M		2024-07
354	Gemini 1.5 Pro (001) Google	53±1	comparison only		2024-05
355	Llama 3.1 Tulu 3 70B DPO Ai2	53±1	local	~22GB	2024-11
356	Qwen Turbo (2024-11-01) Alibaba	53±2	comparison only		2024-11
357	Qwen2.5-VL 32B Instruct Alibaba	53±2	local	~22GB	2025-03
	more from Ai2
259	Olmo 3.1 32B Think Ai2	58±1	local	~21GB	2025-12
262	Olmo 3 32B Think Ai2	58±1	local	~21GB	2025-11
314	Molmo 2 8B Ai2	55±2	local	~6.8GB	2025-12

see all LLMs · find the ones that fit your machine

$ what it's good at

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #230 of 393 * · 8.2% of its score · 2 tests

Beats Mistral Small 4 (119B-A6.5B), Ministral 3 14B Reasoning 2512, Qwen2.5 Coder 32B Instruct here, though they rank higher overall. Trails Flan-T5 XXL, LFM2.5 1.2B Instruct, Gemma 3 4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AlpacaEval 2 length-controlled lab	49.6% ✎	▁▁▃▄▆██▆▃▃▁▁ beats 127/155	medium · avg 47%	▇▇ 5.4%	at its level · small run
IFEval prompt-level loose lab	82.6% ✎	▁▂▁▂▁▃▁▂█▂▂▁ beats 16/25	easy · avg 79%	▇ 2.8%	small run

math▁▂▃▅▆▅▄█▄▁▁▁ #267 of 398 · 47% of its score · 4 tests

Beats Molmo 2 8B, Llama 3.3 70B Instruct, DeepSeek-V2.5 here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Mixtral 8x7B v0.1 (base), InternVL2-4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MATH level 5	42.7% ✎	▂▂▅▆▄█▆▃▅▂▃▄ beats 35/92	hard · avg 32%	▇▇▇▇▇▇▇▇▇▇▇ 28%	sharp · at its level
MATH (subset unspecified) lab	62.3% ✎	▂▄▄▃█▇▅▆▄▃▄▁ beats 36/80	medium · avg 58%	▇▇▇ 6.7%	at its level · small run
GSM8K lab	93.5% ✎	▂▂▂▅▅▆▅▇█▇▄▃ beats 134/146	very easy · avg 87%	▇▇ 6.2%	sharp · small run
OTIS mock AIME 2024–2025	4.4% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 29/193	very hard · avg 8.5%	▇▇ 5.9%	sharp · noisy

other tests6 not in a task · 45% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond	46.3% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 54/208	hard · avg 40%	▇▇▇▇▇▇▇ 16%	sharp · at its level
BIG-Bench Hard lab	84.8% ✎	▁▁▁▁▂▄▅▅█▃▄▁ beats 118/129	easy · avg 77%	▇▇▇▇ 10%	sharp · small run
DROP lab	74.1% ✎	▅··▅·█▅█▅▃▃▅ beats 10/17	easy · avg 65%	▇▇▇ 8.3%	at its level · small run
MMLU lab	83.3% ✎	▁▁▁▁▁▂▃▅█▃▆▃ beats 149/183	easy · avg 79%	▇▇▇ 6.9%	small run
PopQA lab	46.3% ✎	▁▁▂▃▅▃▂█▂▃·▄ beats 33/36	hard · avg 35%	▇ 2.4%	blunt · at its level · small run
TruthfulQA MC2 lab	67.9% ✎	█▅▇▅▁▄▁▇▅▄▁▁ beats 64/73	medium · avg 64%	▇ 0.9%	blunt · at its level · small run

$ where to run it

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Llama 3.1 Tulu 3 70B DPO · needs ~42.5 GB · from huggingface.co/bartowski
chat in the terminal
llama cli -hf bartowski/Llama-3.1-Tulu-3-70B-DPO-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf bartowski/Llama-3.1-Tulu-3-70B-DPO-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Llama 3.1 Tulu 3 70B DPO · needs ~42.5 GB · from huggingface.co/bartowski
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/bartowski/Llama-3.1-Tulu-3-70B-DPO-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Llama 3.1 Tulu 3 70B DPO · needs ~42.5 GB · from huggingface.co/bartowski
chat in the terminal
ollama run hf.co/bartowski/Llama-3.1-Tulu-3-70B-DPO-GGUF:Q4_K_M

MLX (Mac)

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Llama 3.1 Tulu 3 70B DPO · needs ~39.7 GB, Mac only · from huggingface.co/Scotto2025
chat in the terminal
mlx_lm.chat --model Scotto2025/Llama-3.1-Tulu-3-70B-DPO-Q4-mlx
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model Scotto2025/Llama-3.1-Tulu-3-70B-DPO-Q4-mlx

Other ways to run it

Run it as published · needs ~169.82 GB, NVIDIA only · from huggingface.co/allenai
vllm serve allenai/Llama-3.1-Tulu-3-70B-DPO

Weights and documentation: huggingface.co/allenai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 50%, Tülu 3: Pushing Frontiers in… 49%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	3	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 50%	CC BY 4.0
Tülu 3: Pushing Frontiers in Open Language Model Post-Training (Lambert et al., Ai2) · lab-reported	9	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 49%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 53.1

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json